ARCHNES
Automated, highly scalable RNA-seq analysis
M ASSA HUSETS INS
ITUTE
by
Rory Kirchner
RSEP 24 2015
B.S., Rochester Institute of Technology (1999)
LIBRARIES
Submitted to the Department of Health Sciences and Technology
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in Health Sciences and Technology
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
September 2015
D Massachusetts Institute of Technology 2015. All rights reserved.
Signature redacted
.
Author.
Department of Health S ences and Technology
2015
Septem
Certified by...
Signature redacted
Martha Constantine-Paton
Professor of E rain and Cognitive Science
Thesis Supervisor
Signature redacted
....... .
.
Acrented by . . . . . .
Emery N. Brown
Program in Health Sciences and Technology
Director, HarvardProfessor of Computational Neuroscience and Health Sciences and
Technology
F
Automated, highly scalable RNA-seq analysis
by
Rory Kirchner
Submitted to the Department of Health Sciences and Technology
on September 1, 2015, in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy in Health Sciences and Technology
Abstract
RNA-sequencing is a sensitive method for inferring gene expression and provides additional information regarding splice variants, polymorphisms and novel genes and
isoforms. Using this extra information greatly increases the complexity of an analysis
and prevents novice investigators from analyzing their own data. The first chapter of
this work introduces a solution to this issue. It describes a community-curated, scalable RNA-seq analysis framework for performing differential transcriptome expression, transcriptome assembly, variant and RNA-editing calling. It handles the entire
stack of an analysis, from downloading and installing hundreds of tools, libraries
and genomes to running an analysis that is able to be scaled to handle thousands
of samples simultaneously. It can be run on a local machine, any high performance
cluster or on the cloud and new tools can be plugged in at will. The second chapter
of this work uses this software to examine transcriptome changes in the cortex of a
mouse model of tuberous sclerosis with a neuron-specific knockout of Tscl. We show
that upregulation of the serotonin receptor Htr2c causes aberrant calcium spiking in
the Tscl knockout mouse, and implicate it as a novel therapeutic target for tuberous
sclerosis. The third chapter of this work investigates transcriptome regulation in the
superior colliculus with prolonged eye closure. We show that while the colliculus
undergoes long term anatomical changes with light deprivation, the gene expression
in the colliculus is unchanged, barring a module of genes involved in energy production. We use the gene expression data to resolve a long-standing debate regarding
the expression of dopamine receptors in the superior colliculus and found a striking
segregation of the Drdl and Drd2 dopamine receptors into distinct functional zones.
Thesis Supervisor: Martha Constantine-Paton
2
Title: Professor of Brain and Cognitive Science
3
Acknowledgments
Thank you to the members of the Constantine-Paton Lab, Robert and Ellen Kirchner,
my parents, Dr. Melcher for her unwavering support and to the Imaginary Bridges
Group.
4
Contents
1
1.1
2
11
Automated RNA-seq analysis
Background . . . . . . . ..
. . . . . . . . . . . . . . . . . . . . . . .
.............................
12
12
1.1.1
RNA-seq ........
1.1.2
RNA-seq processing challenges
1.1.3
RNA-seq analysis pipelines
. . . . . . . . . . . . . . . . . . .
15
1.1.4
Scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
1.1.5
Reproducibility . . . . . . . . . . . . . . . . . . . . . . . . . .
23
1.1.6
Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
1.1.7
Q uantifiable . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
1.1.8
RNA-seq implementation
. . . . . . . . . . . . . . . . . . . .
31
1.1.9
Variant calling and RNA editing
. . . . . . . . . . . . . . . .
45
1.2
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
1.3
Future development . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
.................
13
Transcriptome defects in a mouse model of tuberous sclerosis
50
2.1
Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
50
2.1.1
Tuberous sclerosis
. . . . . . . . . . . . . . . . . . . . . . . .
50
2.1.2
Tuberous sclerosis is a tractible autism model . . . . . . . . .
59
5
2.1.3
2.4
3
63
2.2.1
Cortex collection . . . . . . . . . . . . . . . . . .
. . . . . .
63
2.2.2
Library preparation and sequencing . . . . . . . .
. . . . . .
64
2.2.3
Informatics analysis
. . . . . . . . . . . . . . . .
. . . . . .
65
2.2.4
Calcium Imaging . . . . . . . . . . . . . . . . . .
. . . . . .
67
R esults . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . .
68
2.3.1
Differential expression
. . . . . . . . . . . . . . .
. . . . . .
68
2.3.2
Pathway analysis . . . . . . . . . . . . . . . . . .
. . . . . .
75
2.3.3
Isoform differential expression . . . . . . . . . . .
. . . . . .
75
2.3.4
Disease association . . . . . . . . . . . . . . . . .
. . . . . .
76
2.3.5
RNA editing
. . . . . . . . . . . . . . . . . . . .
. . . . . .
80
2.3.6
Syn1-Tsc1-/- mice have aberrant calcium signaling.
. . . . . .
81
2.3.7
Htr2c blocker halts aberrant calcium signaling . .
. . . . . .
83
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . .
83
.
.
.
.
.
.
.
.
.
.
.
. . . . . .
.
2.3
61
Methods . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
2.2
The second order effects of mTor activation are important
Transcriptome independent retained plasticity of the corticocollic87
ular projection of the mouse
3.1
3.2
. . . . . . . . . . . .
87
3.1.1
Superior colliculus . . . . . . . . . . . . . . . . . . . . . . . .
87
3.1.2
Retinotopographic map formation in the colliculus
. . . . . .
91
Background ....
. . ..
.........
. .
M ethods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.2.1
Eye closure manipulation and colliculus dissection . . . . . . . 100
3.2.2
Total RNA isolation and quality control
3.2.3
cDNA Library creation and sequencing . . . . . . . . . . . . . 104
3.2.4
Informatics analysis
. . . . . . . . . . . . 101
. . . . . . . . . . . . . . . . . . . . . . . 105
6
3.3
3.4
3.2.5
Differential expression . . . . . . . . . . . . . . . . . . . . . . 105
3.2.6
Corticollicular projection mapping and quantitation . . . . . . 106
R esults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
. . . . . . . . . . . . 107
3.3.1
Corticocollicular projection remodelling
3.3.2
Sequencing quality control . . . . . . . . . . . . . . . . . . . . 108
3.3.3
Differential gene expression
3.3.4
Possible X-linked cofactors in LHON . . . . . . . . . . . . . . 113
3.3.5
Differential exon expression
3.3.6
Dopamine receptor expression in the superior colliculus . . . . 114
. . . . . . . . . . . . . . . . . . . 108
. . . . . . . . . . . . . . . . . . . 114
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7
List of Figures
1-1
collectl benchmarks of hourly memory, disk and CPU usage during a
RNA-seq analysis.
1-2
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
Schematic of parallelization abstraction provided by ipython-clusterh elp er. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
1-3
The Poisson distribution is overdispersed for RNA-seq count data. . .
30
1-4
RNA-seq differential expression concordance calculation from two simulated experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
1-5
Schematic of RNA-seq analysis. . . . . . . . . . . . . . . . . . . . . .
34
1-6
Tuning adapter trimming with cutadapt
. . . . . . . . . . . . . . . .
36
1-7
Gene level quantification can introduce errors. . . . . . . . . . . . . .
39
1-8
RNA-seq improves the rat transcriptome annotation.
. . . . . . . . .
44
2-1
CNS manifestations of tuberous sclerosis. . . . . . . . . . . . . . . . .
51
2-2
The mTor pathway is affected in many disorders with autism as a
phenotype.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2-3
The laminar structure of cortex in Syn1-Tscl-/- mice is undisturbed.
2-4
MA-plot of gene expression in the cortex of wild type vs. Synl-Tscl-litterm ates.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
54
58
69
. . .
2-5
Culurued Syn1-Tsc1-/- neurons have synchronized calcium flux.
2-6
Blocking the Htr2c receptor halts aberrant calcium spiking in Syn1-
82
Tscl-/- neurons. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
84
3-1
Cell types of the superior colliculus . . . . . . . . . . . . . . . . . . .
90
3-2
Eye opening refines the corticollicular projection.
99
3-3
Schematic of superior colliculus dissection. . . . . . . . . . . . . . . . 102
3-4
Effect of eye closure state on corticocollicular arbor development.
3-5
Trimmed mean of M-values (TMM) normalization effectively normalizes gene expression.
. . . . . . . . . . .
. . 109
. . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3-6
Clustering of gene expression in the rat colliculus. . . . . . . . . . . . 111
3-7
MA-plot (mean expression vs log2 fold change) between eyes open and
closed rats.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3-8
MA-plot of differential exon analysis. . . . . . . . . . . . . . . . . . . 115
3-9
Dopamine receptor subtypes are segregated into distinct layers of the
superior colliculus.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
9
List of Tables
1.1
Cleaning the raw Cufflinks reference-guided transcriptome assembly
improves the false positive rate. . . . . . . . . . . . . . . . . . . . . .
43
2.1
Differentially expressed genes in Synl-Tscl-/~ mice.
. . . . . . . . . .
75
2.2
KEGG pathway analysis of differentially regulated genes . . . . . . .
76
2.3
Gene Ontology (GO) term analysis of differentially regulated genes.
77
2.4
mTor signaling is differentially regulated in the Syn1-Tsc-/- mouse.
78
2.5
Autism, intellectual disability and epilepsy genes differentially regulated in the Syn1-TscEl-/- mouse . . . . . . . . . . . . . . . . . . . . .
79
2.6
RNA editing events found in Synl-Tsc-/- and WT mice.
80
3.1
Power estimation of the eyes open vs. eyes closed experiment at a
. . . . . . .
range of fold changes . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.2
Differentially expressed genes with FDR < 0.05 . . . . . . . . . . . . 113
Chapter 1
Automated RNA-seq
analysis
RNA-sequencing (RNA-seq) has largely supplanted array-based methods for inferring differential expression at the gene, isoform and event level as RNA-seq is more
sensitive and produces more information than array-based methods at a similar cost.
However processing RNA-seq data and making use of the extra information that
comes from having sequence data is complex and experiment dependent. Furthermore what is considered best practice is constantly changing which presents a researcher with many choices of how to perform each step of an analysis. Analyzing
a large RNA-seq dataset requires access to a high performance computing cluster,
either locally or via the cloud. Even a moderately sized RNA-seq experiment can be
hundreds of gigabytes in size and if care is not taken with the choices of tools, they
may not operate properly on data of this scale. This chapter describes the RNAseq module implemented for this thesis as part of the bcbio-nextgen project, a
1.1. BACKGROUND
12
community-developed, highly scalable and easily installable set of analyses of whole
genome and exome variant calling and RNA-seq data.
This software is in use in
academic research labs, core facilities and pharmaceutical companies all over the
world, is downloaded thousands of times every month and has processed thousands
of RNA-seq samples for hundreds of researchers all over the world[37].
1.1
1.1.1
Background
RNA-seq
RNA-sequencing has largely supplanted array-based methods for quantifying geneexpression. Array based methods estimate gene expression by attaching small antisense probe sequences designed to bind a subset of the transcriptome to a slide.
Query sequences are fragmented, labelled with a fluoroescent dye and hybridized
to the array.
The luminance at each spot on the array is measured and used as
a proxy for the expression of the gene the probe is designed to target, assuming a
one to one relationship between the probe and the target gene[125]. This approach
has some limitations. The first limitation is that only known genes can be assayed,
because the probe sequences must be designed from known transcripts. The second
is that because the quantitation is based on luminance, the dynamic range of the
measurements is constrained to just a few orders of magnitude, as at one end the
luminance falls below the level of noise and at the other end the camera is saturated.
The third limitation is that, for the most part, only gene expression can be assayed
using this approach.
RNA-seq removes some of the limitations of array analysis.
RNA-seq assays gene expression via the sequencing of small fragments of cDNA up
to hundreds of bases in size called reads, aligning the reads to the genome of interest
1.1. BACKGROUND
and estimating the overall expression by counting the number of reads aligning to
each gene[186]. Since alignment to the genome is not dependent on the state of the
transcriptome annotation it is theoretically possible to infer new genes and isoforms
strictly from alignments of the reads to the genome. As new genome builds are created and new annotations layered on top of them old data can be further mined for
new information by realigning to the new genome build and reanalyzing the gene
expression using the updated annotation. Since quantitation of gene expression is
based on counting the number of reads aligning to a gene, the dynamic range is
limited by only how many reads are sequenced, lending a much higher sensitivity at
the low and high end of the expression spectrum over array based methods. Finally,
by examining differences in the sequence of the reads aligning to the genome, it is
possible to call variants and RNA-editing events from RNA-seq data. The increased
power does not come for free, however. Making use of this extra information makes
the analysis of RNA-seq data much more complex and places more computational
demands on researchers.
1.1.2
RNA-seq processing challenges
There are several major challenges when handling RNA-seq data. RNA-seq data is
very heterogeneous, there are many choices at each stage of performing a RNA-seq
experiment, starting at choices of how to extract and purify and fragment the RNA,
the type of library preparation performed and the type of sequencing to be performed.
Sequencing can be from one end of the RNA fragment (single) or both ends (paired),
with a range of lengths of reads and insert sizes of the RNA fragment between the
pairs. In addition, RNA-seq data can be from one of several strand-specific protocols
where information regarding the strand a gene is on is preserved during sequencing,
13
1.1. BACKGROUND
14
which affects the downstream analysis. Finally there are several types of sequencers
and each type of sequencer produces reads with varying characteristics and error
profiles, all of which need to be taken into account during an analysis. As a result
of these complexities it is difficult to have a simple one-size fits all analysis pipeline
that can handle a wide variety of experiments. Thus, RNA-seq analysis in published
papers tends to be experiment-specific and difficult to reproduce. We have solved this
problem by implementing a flexible analysis pipeline that can handle many different
types of RNA-seq experiments.
In addition to being very heterogeneous, RNA-seq data can be very large and computationally complex which adds to the difficulty in analysis. A single lane of data
can be tens of gigabytes in size and a large scale RNA-seq experiment can encompass
hundreds of lanes or more. Memory and CPU requirements for most programs mean
most analyses will not be able to be run on common lab computers and will have
to be run on either large-scale shared cluster compute environments or cloud-based
compute environments.
These compute environments themselves are very hetero-
geneous, with differing storage backends, shared and non-shared filesystems, ways
to distribute jobs to a cluster via cluster schedulers, heterogeneous CPU and memory availability on nodes, operating systems and other complexities relating to the
hardware the analysis is run on. Not only must an analysis pipline be able to run in
varied compute environments, it has to be able to process data across a wide range of
scales from single samples to thousands of samples. We have solved these problems
in bcbio-nextgen by abstracting the compute environment away from the analysis
with a small library called ipython-cluster-helper. The ipython-cluster-helper
library has been used in several other projects to provide scaling across all available
high performance computing schedulers[129][128).
Another area of challenge is the large choice of tools for performing each step
1.1. BACKGROUND
15
in an analysis. There are a myriad of programs available to perform every step in
an analysis, from quality control of the raw reads, aligning[54], assembly[58][163]
and quantification[46] of transcripts, differential expression[144] and final quality
control [45] [62] of the results. Often times it is not clear what the tradeoffs are when
choosing one tool over another, and some tools may not function properly with large
datasets or with data from a particular sequencer or type of experiment. Each tool
for each step in an analysis is configurable and determining how to properly configure
each tool for a specific analysis is complex, time consuming and prone to error. For
some tools it can be difficult to even install the tool, especially on shared compute
environments where necessary libraries may not be installed or administrator privileges might be needed to install some programs.
In addition each tool may need
special data files in a wide variety of formats to work properly and producing these
data files from the known genome and transcriptome can be a challenging task.
All of these challenges make it extremely difficult for a researcher without programming experience and without a deep knowledge of UNIX to even get started
analyzing their own data. In addition, all of these challenges make performing analyses in a reproducible way very difficult; at the end of an experiment it is often
impossible to trace what was run on each sample much less reproduce the analysis
using the same tools with the same data.
1.1.3
RNA-seq analysis pipelines
As of July, 2015 there are 200,000 Google results for 'RNA-seq analysis pipeline' but
very few of those are under active development and have kept up with the state of
the art and none of them install everything you need to run a RNA-seq analysis. A
poll on the RNA-seq blog in July 13th, 2013 asked 'What is the greatest immediate
1.1. BACKGROUND
16
need facing the RNA-seq community?' with ~ 70 % of the responses being either a
need for a standardized data analysis pipeline or more skilled bioinformatics experts
to process the data. The rate of generation of sequencing data is far outpacing the
generation of skilled bioinformaticians so it will be important to have an accurate,
flexible, validated RNA-seq pipeline that can be run by a naive researcher in order
to keep up with the rapidly increasing amount of generated sequencing data. The
need for this is reflected in industry; in December, 2014 Bina, a company that has
implemented a variant calling and simple RNA-seq analysis pipeline was bought by
Roche and a myriad of startups have sprung up, promising to handle the analysis
of second generation sequencing data either on local appliances or the cloud. It is
important for a free, community developed, open source alternative to these industry
pipelines to exist because many academic labs will not have funding to run on these
commercial appliances.
It is important to specify what characteristics a standardized, scalable and reproducible RNA-seq analysis pipeline should have. The first is that it should be
configurable at a high level of abstraction.
A proper pipeline should translate a
small set of high level options to the appropriate low-level settings for each tool in
a pipeline.
These exposed options should be a minimal set to accurately run an
analysis, with as many of the low-level options as possible derived from the data. A
researcher should not have to be familiar with the intricacies of each tool in order
to run an analysis. The second characteristic of a standardized RNA-seq analysis
pipeline is it should be reproducible. A full accounting of the versions of all third
party tools that were used along with all of the commands that were run should
be provided at the end of an analysis so the analysis can be reproduced. Ideally
a Docker container with all of the relevant software installed would be distributed
along with the raw data so any researcher could reproduce the analysis. The third
1.1. BACKGROUND
important property of the pipeline is that it should be scalable. The pipeline should
be able to be run on small pilot experiments of a couple of samples and be able
to be scaled up to a full experiment with thousands of samples, assuming there is
sufficient compute to run the analysis. Running on the cloud is a requirement for
a generally useful analysis pipeline, because a research can scale up or down their
compute depending on the size of their analysis. The fourth important property is
that the analysis pipeline should be open and flexible. The pipeline should not be a
black box, and it should be able to be changed easily to keep up with new innovations. When choices are made for each step in an analysis, the reasoning should be
laid out in a document so a researcher can understand why the choices were made.
The fifth important property is that the pipeline should be validated. There should
be a set or sets of independent benchmark data against which changes to the pipeline
are tested, to catch regressions in new versions of the pipeline and to measure what
is an optimal way to process a dataset. Finally the pipeline should produce output
data in an easily understandable, parseable format in the form of a report and well
structured output data for dissemination with downstream informaticians and bench
researchers.
In this work we implemented three tools bcbio-nextgen, bcbio.rnaseq and
ipython-cluster-helper which together produce a RNA-seq pipeline with each of
these characteristics. It has been used to process thousands of samples in dozens of
laboratories all over the world.
1.1.4
Scalability
An analysis is scalable if it can run across a broad range of sizes of experiments, from a
small pilot experiment to experiments with thousands of samples. The issue of scaling
17
1.1. BACKGROUND
18
up is a complex one involving many areas of possible difficulty. At the most basic
level, many third party tools that work on small experiments fail when confronted
with a large number of samples or extremely large datasets. We solve this issue by
either providing tools known to scale well or providing fixes upstream to existing tools
to enhance scalability.
An example of the former is replacing the commonly used
samtools suite of tools used for manipulating BAM files with a high performance
alternative sambamba where possible. An example of the latter is providing parallel
implementations of previously existing tools such as GEMINI[128].
Other areas
where scalability can break down is in overwhelming the I/O of the machine; we
choose algorithms which minimize the amount of I/O they perform and stream as
much as possible between steps in the analysis to improve performance. Members
of the bcbio-nextgen community have implemented[36] support for collection of
performance data in terms of disk I/O and memory, network and CPU usage at each
stage in the analysis using collectl[154]. Figure 1-1 on the following page shows an
example of the usefulness of collecting these metrics when writing a scalable analysis.
We provide scalability across compute architectures by using the IPython.parallel[133]
framework for parallel and distributed computing.
We developed the ipython-
cluster-helper library to provide a set of abstractions on top of cluster schedulers
much like the Distributed Resource Management Application API (DRMMA). DRMMA provides a unified API for submitting jobs to an array of cluster schedulers but
has the limitation that you must have a scheduling system installed on the compute
environment. ipython-cluster-helper expands the possible compute configurations
by allowing machines to work as an ad-hoc cluster with no cluster scheduler, as long
as the machines can be accessed via secure shell (SSH). ipython-cluster-helper
provides a view to a cluster of machines which uses a simple unified interface to
distribute jobs to the cluster of machines. This allows our pipeline to distribute jobs
19
1.1. BACKGROUND
1.1. BACKGROUND
19
000-00 0#
N~
40
E
251i
F
Tf~
2
U
~ThJ~
3
4
a
5
140
120
so
so
20
I
0
-
2
-
- ---
-
I i
4
S
5
a
_
*4$~~_
_
00
AT~
20
2
3
4
5
Figure 1-1: collectl benchmarks of memory, disk and CPU usage during a RNA-seq analysis.
top) memory (in gigabytes), middle) disk (in kilosectors/second, yellow is reads, purple
is writes) and bottom) CPU (in percent CPU usage). Having benchmark statistics for
each step in the analysis helps when optimizing performance. For example, the period
of low CPU usage in this run during expression was due to DEXseq running serially;
fixing this and running DEXseq in parallel cut the time to estimate transcript expression
in half. Similarly the disk intensive portions during trimming were eliminated by moving
to streaming between the trimming steps, (see Figure 1-6 on page 36).
1.1.
20
BACKGROUND
across a wide array of compute architectures; running on a laptop in local mode or
running across a cluster with thousands of computers both use the same interface
to distribute work. Figure 1-2 shows a schematic of the IPython.parallel architecture that ipython-cluster-helper uses for performing parallel computations.
An
IPython cluster is spun up consisting of a controller process and a set of IPython
engines running on arbitrary nodes of a compute architecture. The controller acts as
a scheduler, distributing jobs to be run on worker engines. Communication between
the controller and the engines happens over ZeroMQ message queues; the controller
places a job to be run in the form of a serialized Python function onto a ZeroMQ
queue and the worker unserializes the function, runs it locally, saves the result and
places a serialized version of the result back on the message queue.
.
.ole
ZMQ1
enlgin~e-1
ZMQ2
engin11e-2
Figure 1-2: Schematic of parallelization abstraction provided by ipython-cluster-helper.
For each step in the analysis an IPython.parallel cluster is set up using ipython-clusterhelper, consisting of a controller process which distributes the computation and a set of
worker processes called engines which perform the computation. The controller is controlled
via a client process, in this schematic bcbio. Communication between the client, the
controller and the engines occurs over ZeroMQ message queues and require only that a port
be open for two-way commmunication between the components of the IPython.parallel
cluster. This allows for the compute layer to be abstracted away so that bcbio-nextgen
can work with a simple interface to the compute.
ipython-cluster-helper is not limited to use in bcbio-nextgen, it can be used
to run any Python program in parallel across all cluster schedulers.
Below is an
1.1. BACKGROUND
21
example Python code listing showing ease of taking a simple non-parallel program
and parallelizing it using ipython-cluster-helper.
This library has been used in
other projects that required scaling across multiple machines[128].
Listing 1.1: Parallelizing a serial function with ipython-cluster-helper.
from clusterhelper.cluster import clusterview
from yourmodule import longrunning_function
import sys
if __name__
==
main__:
# serial version
for f in sys.argv[1:]:
longrunning_function(f)
# parallel version with ipython-cluster-helper
with clusterview(scheduler="lsf ", queue="hsph",
num_jobs=5) as view:
view.map(longrunning_function, sys.argv[1:])
All of the scalability in the world is useless if the analysis infrastructure is not able
to be relocated to the compute. Thus having a highly scalable pipeline also requires
having a pipeline that is easily relocatable and that means having an installation
process which installs everything that is necessary for an analysis.
This includes
all of the necessary data including the genome sequences and metadata about the
genome sequences including annotation of gene boundaries and other genomic features of interest. It also involves installing the correct versions of hundreds of tools,
libraries and other programs to run the analysis. These tools have to be installed
whether or not the user has administrative access to the machine as many researchers
using shared compute environments will not have administrative access. We handle
1.1. BACKGROUND
installation by contributing back to two major projects for installing bioinformatics
tools, the cloudbiolinux[2] project and the Homebrew project. Both projects provide simple to specify formula for getting and installing tools on UNIX and MacOS
machines. bcbio-nextgen uses these two projects to compile and install the tools
necessary to run an analysis.
The combination of abstraction of the underlying compute with ipython-clusterhelper and the ability to install all of the necessary tools in an environment-agnostic
way allows bcbio-nextgen to be able to easily installed and run in a wide variety
of compute environments. Recent work on bcbio-nextgen by Brad Chapman and
John Morrissey, in collaboration with groups at Biogen, AstraZeneca and Intel have
implemented two extremely useful scalability features in bcbio-nextgen. The first
is enabling bcbio-nextgen to be run using the container system Docker[8]. Rather
than running the installation script to install all of the tools, installation requires
just downloading a Docker container with bcbio-nextgen and all of the dependencies
already installed. The second scalability feature is a set of tools to use the Docker
images to run an analysis on Amazon Web Services (AWS), pushing and pulling
data from Amazon's storage solution, S3. This is an example of one of the major
benefits of working on an open source, community driven project: infrastructure like
Docker and AWS support only needs to be implemented once for all analyses to take
advantage of it. Since the analysis is abstracted away from the compute architecture
it also means rerunning an old analysis can be done on more modern hardware or by
other researchers to reproduce an old result.
22
1.1. BACKGROUND
1.1.5
23
Reproducibility
High throughput genomics has come under scrutiny in the past couple of years for
having issues with reproducibility. Here, our definition of reproducibility is broad; we
do not mean reproducing the exact numerical result of an analysis but reproducing
the major finding of a paper.
Even with this relaxed definition, many genomics
experiments cannot be reproduced. This problem is so widespread that the NIH has
undertaken an initiative to enhance reproducibility in genomics[40].
One pressing
issue is there are often no standards for experimental design or analysis[104], and
this is especially true for gene expression experiments. The lack of community derived
standards results in widespread reproducibility problems with both microarray[50]
and RNA-seq experiments[124].
For an informatics analysis to be reproducible, more than just the data has to
be made available. Complete recounting of the command lines of all software, the
versions of all software, the versions of all intermediate plumbing type code and code
generating the downstream summary statistics must be provided.
Not only must
this software be made available, it must have to work in the compute environment
of the person reproducing the analysis. To be reproducible, the code generating the
results must be open; a description of the algorithm is not sufficient to reproduce
the results of the code[77].
We address this problem of reproducibility in several
ways. The first is that we install all third party tools needed to run an analysis on a
wide variety of computing platforms. This allows a person attempting to reproduce
results to start from the same base environment. The second is that we record all
versions of all third party tools and all command lines run so that an analysis can be
reproduced by running the commands if necessary. The third is that if an analysis is
run via Docker, a Docker container that contains everything necessary to reproduce
1.1. BACKGROUND
an analysis pre-installed which can be run on a local cluster computer or on Amazon.
Each version of a container is specified by a unique id which can be accessed later,
meaning an analysis can be completely reproduced by any researcher able to use
Docket containers at a later time. The Docker and'AWS implementation in bcbionextgen further expands reproducibility, since with that integration a researcher
only needs to pay Amazon for the compute and storage and can rerun a published
analysis on their own. In these ways, bcbio-nextgen provides a solution to the
reproducibility problem.
1.1.6
Configuration
RNA-seq analyses are complex in part due to a wide array of choices that one can
make regarding how the RNA is extracted, how the libraries are prepared, how
the sequencing is performed, what organisms are being studied, if the libraries are
stranded or not or aimed at tagging a specific portion of a transcript instead of the
entire transcript. Each of these options has an effect on the downstream analysis. For
example a stranded experiment must take the strand information into account when
aligning, quantifying and assembling the reads, so each tool that is run must have the
appropriate options set. We provide a carefully considered set of high level options
that describe a RNA-seq experiment that an experimenter must set with many of
the low level configuration options for each tool set to appropriate default values or
appropriate values learned from the data. In this manner we simplify setting up a
RNA-seq analysis down to setting a minimum set of parameters that can drive a
wide variety of tools underyling the analysis.
These parameters describe whether the library is strand specific or not, which kit
was used to produce the library, and which genome the library was from. Parameters
24
1.1. BACKGROUND
controlling expensive, optional analyses such as whether or not to call variants or
assemble the transcriptome can be enabled if those were part of the design of the
experiment. We also handle a difficult type of experiment that is often used in cancer
research, where a tumor from one organism is grown in another organism, often a
human tumor in mouse tissue. When the tumor is sequenced, some mouse tissue can
be included. There is an option to disambiguate reads of two organisms in order to
remove this type of contamination. Finally, arbitrary metadata about each sample
can be provided, such as which batch it is from, if it is treated or not, anything the
user can provide. During downstream analyses, this metadata is automatically made
available for model fitting and differential expression calling in a separate RNA-seq
reporting tool we created called bcbio.rnaseq.
1.1.7
Quantifiable
It is important to have validation datasets for an analysis, both to benchmark the
speed of the analysis as well as validate the results of the analysis. Validation datasets
allow for an analysis to be fine tuned in terms of new tools or configuration of
existing tools to improve results without fear of introducing errors into the data.
More importantly it also places the downstream results in a greater experimental
context.
A validation data set allows a researcher to understand where they are
making mistakes or where the analysis is blind and can help guide downstream users
of the data towards the most salient results.
Having a standardized, quantifiable analysis allows a researcher to treat their
analysis as an optimization problem and improve it. Benchmark data sets serve as
type of integration test for the software making up an analysis pipeline. When a new
tool is published, it is standard for the work to show a comparison to the standard
25
1.1. BACKGROUND
tools of the day and show how the new tool outperforms the old in some aspect.
Occasionally review papers compare several tools to each other with a validation
dataset to determine in a less biased way the strengths and weaknesses of each
tool[98][144]. These papers are valuable but often there has been no fine tuning of
any of the comparison algorithms. Consequently, they are often run with their default
values and are compared against each other, whereas a researcher with experience
with a particular tool would be able to tune it to get better results. bcbio-nextgen
is tool agnostic and tools have gone through a process of fine tuning the configuration
parameters already. Tools are chosen based on the consensus of the users regarding
which tool works best and the tools are tuned or improved to give the best results.
This allows for a fair comparison across tools when evaluating making changes to an
analysis pipeline.
In addition to being tool agnostic, bebio-nextgen is also dataset agnostic. If
a new benchmark dataset comes out for a particular aspect of a pipeline, the correct values for the benchmark dataset can be added and the benchmark dataset
run through the pipeline and compared. In this way researchers can improve both
the analysis itself and the measurement of the analysis simultaneously through the
incorporation of new benchmark datasets.
Researchers can ask many more questions of their RNA-seq data than with microarray analyses. Expression can be summarized and differential expression called
at the gene, transcript, exon and splicing event level. Rearrangements and variants,
gene fusion events and RNA-editing events can all be assayed with RNA-seq. Novel
genes and novel isoforms can be assembled, and differential usage of promoter sites
can be analyzed. Each of these aspects of an analysis have several tools which handle
them and each needs a benchmark dataset to test against when testing iterations of
a pipeline. bcbio-nextgen gathers available benchmark datasets from the commu-
26
1.1. BACKGROUND
27
nity to test each aspect of RNA-seq analysis. We provide benchmark datasets in the
form of data from actual experiments and simulated data to benchmark gene-level
and transcript-level expression calling, transcriptome assembly with and without a
reference data set and fusion gene calling on cancer datasets. As the community
develops alternate benchmarks we fold them into bcbio-nextgen.
Gene and transcript expression
bcbio-nextgen uses a combination of validation datasets from real data and simulation to assess the results of the gene-wise differential expression analysis. The first
validation dataset is from phase three of the Sequencing Quality Control (SEQC) [41]
project from the US Food and Drug Administration.
The goal of releasing these
datasets was to provide laboratories with known reference standards to use to tease
out the types of technical variation introduced across laboratories, sequencers and
protocols. This dataset consists of a two sets of samples: the first is RNA sequenced
from the Universal Human Reference RNA (UHRR) from Agilent. This sample is
composed of total RNA from ten human cell lines to be used as a reference panel.
The second sample is from the Human Brain Reference RNA (HBRR) panel from
Ambion which consists of brain samples from all regions in several subjects pooled
together to form a reference panel. The SEQC project provides qPCR results from
one thousand genes in the UHRR and HBRR panels, to serve as a proxy for a truth
dataset. The validation dataset we analyze is fifteen million randomly selected reads
from five replicates of each SEQC sample.
The SEQC data have a few major limitations as a validation data set. The first
limitation is that since the RNA comes from stock of pooled RNA, the replicates are
technical replicates of only the library preparation and sequencing and do not take
1.1.
BACKGROUND
28
into account variability induced by RNA-extraction or biological variability such as
samples from different mice or humans.
Proper handling of biological variability
is important component in a RNA-seq analysis and the SEQC dataset is not an
appropriate way to assess it. The second major limitation is that the validation data
is gene-level qPCR data. RNA-seq data is capable of assessing expression at the level
of the transcript but the SEQC data set will not be useful for assessing quantitation
at anything other than the level of the gene.
To address these limitations the we also produce simulated datasets of raw RNAseq reads using Flux Simulator [67], an in-silico RNA-seq experiment simulator. Using
Flux Simulator, reads are generated from a given reference transcriptome and many
steps and biases introduced during RNA-seq library preparation are simulated including fragmentation at the RNA or cDNA level, reverse transcription, PCR amplification, gene expression and sequencing. Flux Simulator does not produce biological
replicates, so we implemented a simulator to add biological noise and fold change
spike ins to the data generated by Flux Simulator to mimic biological replicates
and differential expression in a real experiment. This package is available online as
flux-replicates.
RNA-seq count and read simulator
Flux Simulator outputs a simulated relative expression level, p which is the proportion of the total RNA molecules in the simulated sample that are from a given
transcript. These values are used as a baseline level of proportions from which a single factor differential expression experiment is simulated by spiking in fold changes
between replicates of a sample. We implemented a RNA-seq read count simulator[94]
by setting a proportion of the simulated transcripts to be differentially expressed at
1.1.
BACKGROUND
29
range of fold changes and drawing counts for each gene from the negative binomial
distribution with the dispersion parameter set to increase as the expression level
decreases, to mimic real data(Figure 1-3 on the following page).
Starting from a given level of biological coefficient variation, BCV, a relative
expression level for gene y with proportion of the total expression level py and a
library size 1for a sample, we simulate cy, the counts of transcript y from the negative
binomial distribution by composing a gamma distribution and a poisson distribution
to add biological and technical noise, respectively:
Py = Py
BCVy = (BCV+
I ) Vx) 2 (4 0)
CY ~ T(B(pty, BCVy))
(1.1)
(1.2)
(1.3)
B is a function that draws values from a gamma distribution of mean Py and variance BCVY 2 and T draws values from a poisson distribution with mean B(py, BCVy).
We add random noise dependent on expression level of the gene to BCV to simulate
the higher dispersion of the negative binomial at low counts (figure 1-3 on the next
page).
The simulator can be run in a mode where it is supplied with an existing set
of counts from which the parameters for the simulation are estimated to match
the existing data.
This allows the user to take a previously run experiment and
estimate what would happen if they sequenced less reads or changed the number of
replicates. It also allows for some rough post-hoc power calculations to inform future
experimental design choices and to make stronger negative result claims rather than
failing to reject the null hypothesis of no differential expression.
1.1. BACKGROUND
30
*
0
I 00000000
C
1000000-
gO
0
1004
7-
Sb
*060;
10
1000
100000
mean count
Figure 1-3: The Poisson distribution is overdispersed for RNA-seq count data. Variance
vs. mean counts for a RNA-seq experiment (black dots) with variance estimated by the
Poisson (black line) and negative binomial (blue line) distributions. Each point on this
graph is the mean and variance for a single gene across biological replicates. RNA-seq data
is more noisy than expected from the Poisson due to additional technical and biological
variability.
1.1. BACKGROUND
As an example, RNA sequencing is very expensive and time consuming and often
experimenters are looking to maximize the power of their experiment and minimize
the cost. One potential optimization involves determining how many reads to sequence against how many replicates to run. An experiment running at most on two
HiSeq lanes can expect at least 300 million reads. This leaves a case-control experiment with three replicates at 50 million reads per sample. Alternatively, rather than
sequencing six samples at 50 million reads per sample, the read depth could be cut
in half and twice as many replicates could be sequenced. Using the simulator to
compare these two experiments against each other, we can make a recommendation
to run more replicates at lower depth if the goal is to identify differentially expressed
genes at moderate-to-high fold change (Figure 1-4 on the following page). In addition, if the genes of interest are at the low end of the differential expression spectrum,
we could recommend to not run the experimental at all.
1.1.8
RNA-seq implementation
Figure 1-5 on page 34 shows a high level overview of the RNA-seq pipeline implemented in bcbio-nextgen version 0.8.5a. The implementation covers quality control
of the data, adapter sequence removal, aligning to the genomes, quantifying at the
gene, isoform and exon level, calling variants and RNA-editing events, transcriptome
assembly, classification and filtering and calling differential expression events at the
gene, isoform and exon level using several commonly used tools. There are a multiplicity of third party tools that could be used to perform each step and tools are
chosen based on considering their accuracy, licensing, scalability in terms of CPU, IO
or memory bottlenecks and how well they are actively maintained. When compute
constraints may be a roadblock, alternatives are provided; for example the STAR[48]
31
1.1.
32
32
BACKGROUND
1.1. BACKGROUND
.1to
1.05
4W-~
_h
I
4OO-
zoo-
200
IIII
--
--
edgeR
Voomkimma
400-1
edgeR
voomlmm
-L-
i
0
(a)
(b)
Figure 1-4: RNA-seq differential expression concordance calculation from two simulated
experiments. An RNA-seq experiment was simulated with a sample size of three and a
library depth of 50 million reads (a) and a sample size of six and a library size of 25 million
reads (b). We simulated 200 genes as differentially expressed up or down across each of five
fold changes: 1.05, 1.10, 1.50 and 4.0. Each facet of the graph calculates concordance, false
positive an false negative rate, considering only genes at the specified fold change or greater.
Four commonly-used negative binomial based differential expression callers (DESeq2[101],
DESeq[6], edgeR[149] and limma[94]) were run on the data and their concordance with
the known fold changes from the spike ins were simulated. At large fold changes most
of the spike in genes are correctly called differentially expressed in both experiments but
at fold changes below four, the experiment with a higher number of replicates calls more
concordant differentially expressed genes.
1.1. BACKGROUND
aligner, while very fast and accurate has a large memory requirement which may
make it not feasible to use in some compute environments. In that instance a slower,
less memory intensive alternative in Tophat2[170] is offered (see Figure 1-5, alignment box).
In the following sections a brief, non-comprehensive rationale behind
choosing the tools for each step in the analysis is covered. Where appropriate there
is a discussion of optimizations we implemented to improve the accuracy and speed
of the chosen tools. This discussion isn't intended to be an exhaustive comparison
of every tool for each step or a conclusive statement about which tools are better
than others, but more a breakdown of the tradeoffs, possible pitfalls and other important considerations for each step in the analysis. The design philosophy of the
bcbio-nextgen RNA-seq pipeline implementation is to use the data to tune as many
parameters as possible to optimize for the particular data set being processed while
maintaining a balance between accuracy and throughput and the tool choices reflect
that philosophy.
Read trimming
RNA-seq reads can have contaminating sequences at the ends of the reads in the form
of adapter sequences, poly-A tails or other sequences, which may cause issues during
alignment or in downstream analyses, resulting in loss of information. Some splicedread aligners such as STAR will handle these reads by soft clipping the homopolymer
or adapter sequences that don't align to the genome but other aligners such as Tophat
do not handle these reads well. For RNA-seq, adapter contamination on the ends of
reads can often be a problem; under most library preparation protocols the RNA is
fragmented before conversion to cDNA and there may be small pieces of RNA for
which the read length is longer than the fragment. For these small RNAs with long
33
34
34
1.1. BACKGROUND
1.1. BACKGROUND
bcbio-nextgen_
trimming
quality control
FastQC
QualiMap
bamstats
cutadapt
p
V~K
assembly
Cufflinks
4
STAR
Oncofuse
genome alignment
STAR (high memory)
Tophat2
transcriptome alignment
STAR (high memory)
bwa-mem
filter/annotate
CPAT
Cuffcompare
fusion gene
gene DE
RSEM
Cufflinks
DESeq2
edgeR
voom + limma
transcript DE
gene expression
featureCounts
eXpress
Cufflinks
transcript expression
DEXseq
RSEM
DEXseq
eXpress
Cufflinks
Figure 1-5: Schematic of RNA-seq analysis. A high level overview of the implementation
of the RNA-seq pipeline in bcbio-nextgen version 0.8.5a, with the external programs
used in each module listed. Included are modules to do gene, isoform and exon-level
differential expression calling, variant calling on RNA-seq data, fusion gene calling for
assaying structural variation in cancers, transcriptome assembly, quality control of the
raw data and the alignments, clustering and sample outlier detection and automatic report
generation. Red colored boxes indicate user-supplied data and green boxes are functionality
supplied by bcbio.rnaseq.
1.1. BACKGROUND
35
read lengths, the read will be long enough to continue on past the RNA sequence
into the adapter sequence. Several tools that have been created to fix this adapter
read through issue, with tradeoffs regarding specificity, sensitivity and speed[83].
Careful tuning of the parameters of the read trimmer beyond the default values
can improve sensitivity and specificity. Figure
1-6a on the following page shows
that tuning cutadapt[109] to trim twice on the forward reads, once to trim possible
adapter read-through and once to trim polyA sequences that are masked by the
adapter sequence can rescue reads that would have been lost from the 3' end of RNAs.
Beyond tuning for accuracy, paired-end read trimming using cutadapt can be tuned
to run faster on compute environments with spinning disks by a simple architecture
change. When run in paired-end mode, cutadapt must write two temporary files to
disk and be run twice, this I/O operation is very slow and takes time. If running
thousands of samples simulataneously, this can become a large bottleneck in an
analysis. If instead of writing temporary files to disk, the command is constructed
to use named pipes[164] instead, a 30% gain in performance can be realized, even on
a very small dataset of a single lane of a million reads.
There has been some debate regarding if it is useful to trim low quality ends
of reads before aligning. Most, but not all, modern aligners can take into account
the low quality scores so most analysis pipelines include a quality trimming step.
The commonly used threshold for RNA-seq is to to trim bases with PHRED quality
scores less than 20, but a more gentle threshold of 5 results in much less information
lost[105].
1.1.
36
36
BACKGROUND
1.1. BACKGROUND
adapter
poly-A
target
fw
fw once
fw twice
rV
rv once
(a)
hard drive
streaming
time (seconds)
SSD
no
61
SSD
yes
64
platter
no
211
platter
yes
154
(b)
Figure 1-6: Tuning of adapter trimming with cutadapt. a) Trimming of adapter, polyA
tails and other non-informative contaminant sequences from the ends of reads is necessary
for compatibility with downstream tools that cannot handle reads with contaminated ends,
such as aligners that do no soft clipping or kmer counting algorithms. cutadapt needs
the flag set to try trimming twice when handling RNA-seq data to be able to trim polyA
tails masked by adapter sequence. b) cutadapt requires intermediate files to be written
out when handling paired end data and cannot natively stream the files from one step to
another. For small amounts of data on fast disks (SSDs), this does not contribute to the
processing time at all. On slow disks writing the intermediate files is the bottleneck in the
process. Replacing the intermediate files with named pipes, thus re-enabling streaming,
speeds up processing by 30% on a million paired-end reads.
1.1. BACKGROUND
Alignment
RNA-seq reads may cross exon-exon boundaries and when aligned to the genome
will appear to be split across an exon. RNA-seq aligners have to be split-read aware
or be able to take a gene model and create a proper alignment for these reads.
The STAR aligner[48] is very fast and accurate aligner for RNA-seq that can map
reads up to 50 times faster than Tophat2 but requires a machine with 50 GB of
memory to run[54]. In addition to mapping reads quickly and accurately, STAR can
simultaneously generate a mapping to the transcriptome for use with downstream
quantitation tools such as eXpress[146], saving a step in the downstream analysis.
When high-memory compute is not available, Tophat2 is run instead of STAR. For
paired-end reads, a small subset of the reads are mapped with Bowtie2 to determine
an estimate of the mean and standard deviation, using the median and median
absolute deviation as proxies for the mean and standard deviation of the insert size.
The transcriptome-only mapping is made using bwa-mem since there is not a need
to handle the intron spanning reads and bwa-mem is extremely fast and sensitive.
Transcriptome expression quantification
The primary goal of most RNA-seq experiments is to examine differences in the
transcriptome between two or more experimental conditions. In order to assay the
differences in transcription, the transcriptome must first be quantified. Depending
on the organism, quantifying the expression of the transcriptome can be more complex than it initially seems. In organisms that have little to no alternative splicing
of transcripts, this task is conceptually simple: count up the number of reads mapping to each gene and determine if the number of reads mapping to the gene is
systematically different between conditions. For organisms such as the human with
37
1.1. BACKGROUND
transcriptomes that undergo intensive alternative splicing[181], the task of calling differences between conditions is much more complex. The complexity arises because
what is sequenced are small pieces of transcripts which, when alternatively spliced,
can be very similar to each other. Figure 1-7 on the next page shows an example
where a single gene has multiple transcripts. Determining which isoform to assign
reads that could come from multiple transcripts is a complex chicken-and-egg problem since it requires knowledge of how much of each transcript is expressed, which
is what you are trying to estimate.
Many approaches ignore the read assignment issue altogether by quantifying at
the level of the gene; the reads aligning to all transcripts of a gene are counted
and combined to be the total reads mapping to the gene. The trouble with this
approach is it doesn't reflect reality, the reads came from individual transcripts, not
a gene and quantifying expression in this way can lead to errors. Figure 1-7 on the
following page shows an example of one type of error, where a splicing event results
in the expression of a much smaller transcript in one condition resulting in a false
differential expression call.
Other practical concerns make quantifying the expression of the transcriptome
difficult.
Quantifying transcripts is dependent on the state of the annotation of
the transcriptome because isoforms of genes that do not occur in the transcriptome
will not have reads assigned to them, and reads that were sampled from unannoted
isoforms will be misattributed to related isoforms, an issue illustrated in cartoon
form in Figure 1-7 on the next page. The state of the transcriptome annotation for
model organisms varies widely even for closely related organisms that should show
similar degrees of alternative splicing (e.g. Figure 1-8 on page 44). Thus, for most
organisms, augmenting the existing transcriptome assembly is necessary for accurate
differential isoform calls. However this thesis, (Table 1.1 on page 43) and work of
38
39
1.1. BACKGROUND
39
1.1. BACKGROUND
isoform A
isoform B
(a)
isoform A
isoform B
isoform C
(b)
Figure 1-7: Gene level quantification can introduce errors. a) Quantification at the gene
level can introduce quantification errors. If in one condition isoform A is expressed and
in another condition isoform B is expressed, quantification at the gene level by counting
the number of reads aligning to each gene will show a 1.5x fold change, even though
the expression of the gene is unchanged. b) Missing isoforms can introduce errors in
isoform-level differential expression calls. In the illustration, isoform C is missing from
the annotation which will lead to reads being incorrectly assigned to isoform A. This also
illstrates the complexities of assigning reads to specific isoforms; if a read aligns to the
first exon, determining to which transcript of the gene it should be assigned is not a trivial
problem.
1.1. BACKGROUND
others[3] have shown that transcriptome assembly is often incomplete and rife with
false positives. Including these error prone assemblies introduces a major source of
noise and into quantifying the transcriptome expression.
The challenges in quantifying at the isoform level have lead to the exploration of
algorithms quantitating individual splicing events, exons and parts of exons instead
of isoforms[87][7]. Quantitating at this level requires much less accurate transcriptome information and only requires enumeration of the exons and splicing events
that can occur in the data, a much more tractable problem than a complete enumeration of all possible isoforms. In addition, for incomplete transcriptome annotations,
assembling exons is much more successful than assembling entire isoforms[3], especially for organisms in which splicing is complex, and the enumeration of the exons is
generally more complete in exisiting transcriptome annotations. Recently a method
called derfinder[60] was developed which takes the resolution of the transcriptome
quantitation to the extreme and quantitates transcriptome expression at the level of
a single base.
Transcriptome quantification is implemented at three levels of resolution in bcbionextgen, at the level of the gene with featureCounts[97], eXpress[146] and Cufflinks[172],
the isoform level with eXpress and Cufflinks and the sub-exon level with DEXseq[7].
featureCounts and eXpress both generate gene-level estimated counts of reads mapping to genes, suitable for use in count-based differential expression callers (see Section 1.1.8 on the following page); featureCounts only counts reads which can be
uniquely assigned to a gene whereas eXpress assigns ambiguous reads probabilistically based on the overall expression of the gene. There are other tools with similar
functionality to featureCounts and eXpress, but featureCounts and eXpress are both
extremely fast, up to 30 times faster than similar tools[97]. For isoform expression,
both eXpress and Cufflinks produce estimates of the gene-level and isoform level
4O
1.1. BACKGROUND
41
expression, eXpress summarizing with estimated counts suitable for use in countbased callers and Cufflinks with FPKM, a gene-length normalized expression measure, which a companion program Cuffdiff uses to call differential isoform expression.
Finally DEXseq quantitates differential expression at the level of the exon fragment;
exons are broken into the smallest fragments unique to an isoform in the existing
annotation and DEXseq quantitates the expression of those fragments.
Each of
these levels of quantitation are used to make differential expression calls in a companion tool released with bcbio-nextgen, bcbio.rnaseq, this gives the researcher
flexibility to choose the resolution of quantitation that is most appropriate to their
experiment.
Differential expression
Differential expression calling on the gene, isoform and splicing event level is performed with a companion program to bcbio-nextgen called bcbio.rnaseq. bcbio.rnaseq
runs baySeq[72], DESeq2[101], edgeR[149], Cufflinks[171], voom+limma[94], edgeRun[47]
and EBSeq[95] to call gene-level differential expression, Cufflinks and EBSeq to call
isoform-level differential expression and DEXSeq to call splicing event level differential expression using the estimated expression values calculated from the bcbionextgen RNA-seq pipeline. Running several tools is important, as different tools
perform well on specific types of RNA-seq data. DESeq2 and limma are great choices
for most RNA-seq experiments, but they can be outperformed in specific conditions.
For example for low replicate, low count experiments with under ten million reads
per sample or less, we have created an improved algorithm called edgeRun[47], which
is much more sensitive for these specific types of experiments.
1.1. BACKGROUND
42
Transcriptome augmentation
A strength of RNA-seq over microarray analysis is that novel isoforms, novel genes
and other novel coding or noncoding RNA can be discovered. With a microarray
analysis the gene fragments assayed are generally restricted to the set of what is
already known. There has been a lot of work on how to augment the known transcriptome with novel isoforms and genes, through either de-novo assembly of the
transcriptome or using the known transcriptome as a reference to guide the discovery of new isoforms. This problem is extremely difficult to solve, and the state of
transcriptome assembly software reflects that difficulty. For complex transcriptomes
with many splicing events such as human, only
30% of the full length transcripts
can be recovered from simulated data, with a similarly low precision[163].
The transcriptome assembly implementation in bebio-nextgen accounts for the
difficulties in de-novo assembling the transcriptome by using the reference guided
assembly mode in Cufflinks, which uses the known transcriptome as a guide to call
novel isoforms and novel genes[148]. Even with this crutch, the false positive rate
of Cufflinks assembly is very high[3]. We confirmed this by simulating reads from
chromosome 22 of the human genome using gene models from Ensembl release 75
and assembling the transcriptome with Cufflinks v2.1.1, using an annotation where
half of the known annotated transcripts are dropped at random.
This resulted in
an assembly where nearly half of the novel transcripts are false positives. We implemented a simple filtering algorithm by training a logistical model[183] to predict
coding/non-coding status of the novel transcripts, training against the known transcripts, and keeping isoforms only if the coding status matches the known coding
status of the known gene. This filtering greatly reduces the false positive rate at the
cost of a moderate hit in sensitivity on simulated data(Table 1.1 on the next page).
1.1. BACKGROUND
43
status
specificity
sensitivity
unfiltered
0.51
0.70
filtered
0.85
0.63
Table 1.1: Cleaning the raw Cufflinks reference-guided transcriptome assembly improves
the false positive rate of the assembly. Unfiltered Cufflinks-assembled transcripts have a
sensitivity of 0.7 but a specificity of 0.5, indicating that half of the assembled transcripts
are false positives. Filtering the assembled transcripts by removing transcripts with a
predicted coding status differing from the known coding status of the parent gene in the
annotation improves the specificity to 0.8 at a minor hit to the sensitivity. Reads were
simulated from chr21 of the human genome using gene models from Ensembl release 75
and Cufflinks was run in reference-guided mode with an assembly missing half of the known
annotated transcripts.
Using data from a real experiment in Chapter 3 of this work, we reasoned that
the rat and mouse should have similar numbers of transcripts per gene since the
two species are highly related. The mouse annotation has many more genes per transcript identified, so we assembled the rat transcriptome and compared the number of
transcripts per gene identified before and after filtering. Prior to filtering, more transcripts were assembled from the rat transcriptome, even though the samples should
capture transcripts expressed in the brain, about 50% of the total transcripts in the
organism (see Chapter 3 on page 87). Our filter greatly reduced the total number of
novel transcripts identified, while still increasing the number of transcripts by 20%
(Figure 1-8 on the next page). The coding/noncoding prediction part of the filter
is also used to classify novel genes, tagging novel genes with a low coding score as
ncRNAs or lncRNAs, depending on their length, and genes with a high coding score
as protein coding. The assembly, model training, filtering and classification are run
automatically during the RNA-seq pipeline implemented in bcbio-nextgen.
44
1.1. BACKGROUND
1.1. BACKGROUND
44
1
s
-
40000
35000-4
+
..
.
--.
--
30000
-----
assembled, cleaned
assembled, raw
EV.
-
Ensembl 75 mm10
-
Ensembl 75 rnS
25000 --
20000-
150000
10
10
20
30
40
transcripts per gene
Figure 1-8: RNA-seq improves the rate transcriptome annotation. The rat (rn5) Ensembl
gene annotation lags behind its close relative, the mouse (mm10), annotation in numbers
of transcripts per gene identified. Using RNA-seq data from the rat superior colliculus,
the unfiltered Cufflinks assembly identifies more transcripts than exist in the mouse gene
annotation even though only 50% of the transcripts of the rat are expressed in the superior
colliculus. Filtering the transcripts by keeping only those transcripts that agree with the
coding status of their parent gene greatly reduces the number of novel isoforms identified,
and is a much more reasonable result.
1.1. BACKGROUND
1.1.9
45
Variant calling and RNA editing
Another benefit of sequencing for gene expression analysis over microarrays is the
ability to call variants from the RNA-seq data. The reads are aligned to the genome
and differences in the RNA compared to the reference sequences are called. Calling
variants from RNA-seq data is prone to errors, however.
Variant calling from RNA-seq data is implemented in bcbio-nextgen following
recommendations[78] from the Broad Institute. This is supplemented with custom
filters to remove variants that are likely to be technical artifacts. When aligned to
the genome, RNA-seq reads that cross exon-exon junctions are soft clipped, meaning
those bases are masked during the alignment and HaplotypeCaller, the Broad's variant caller, can't handle the soft clipped bases. We split those reads into two reads,
one for each end of the exon and then call variants with HaplotypeCaller. We then
filter the called variants, throwing out variants that have a large strand bias, as these
are likely to be false positives caused by incorrect mapping. We also drop variants
that are biased in terms of where they appear in a read as these are also likely due
to incorrect mappings.
genome.
Other sources of error come from misaligned reads to the
The transcriptome is much more complex than the genome and a given
exon can be attached to many different exons through alternative splicing. This can
cause many false positive variants to be called where two exons are spliced together
[139]. We filter out variants that are within 5 bp of an exon-exon junction in our final
set of variant calls. During this process we also keep all variants that are known to
exist for that organism using the most current dbSNP release and pass them through
unfiltered.
Variants can also be identified in the RNA sequences not due to variants in the
DNA sequence, but by variants in the RNA sequence introduced by A-to-I RNA-
1.1. BACKGROUND
editing by a class of enzymes called ADARs[152].
46
For supported organisms, we
separate out variants that are likely to be RNA-editing events from DNA variants
using a combination of preexisting annotations and custom filters. The majority of
RNA editing events verified to date are A-to-I edits, where the I is interpreted as
G during translation. For mouse and human samples, possible RNA-editing events
that appear in the the Rigorously Annotated Database of A-to-I Editing (RADAR)
and the Database of RNA Editing in Humans (DARNED) are kept.
Functional effects of the cleaned DNA and RNA-editing events are calculated
using the Variant Effect Predictor from Ensembl. Human calls are then loaded into
a GEMINI[128] database, which loads the variants into a database and annotates
the variants with information from their genomic context, using existing variant
annotations from ENCODE, OMIM, dbSNP, KEGG, HPRD and other databases.
Experiment summary
bcbio.rnaseq produces a summary report of the output from the bcbio-nextgen
pipeline in the Rmarkdown language, a hybrid of the markdown document formatting language and the R statistical analysis language. The report includes a wealth
of information for each sample including the number of reads mapped, the mapping
rate, the number of genes and isoforms detected, the amount of rRNA contamination, the overall complexity of the library, and how well the transcriptome was
covered. Heatmaps of correlations between samples are included in the plot as well
as multidimensional scaling (MDS) plot; these plots along with the basic metrics
regarding each sample are helpful for eliminating outlier samples from the analysis.
If a model formula is included in the bcbio-nextgen run, bcbio.rnaseq will run
the differential expression callers on the output and generate a summary report of
1.2.
DISCUSSION
the differentially expressed genes found in the experiment. The RMarkdown report
serves as a starting point for digging deeper into an experiment. It has all of the
metadata supplied about each sample, the calculated quality metrics, the normalized count data, the differential expression calls and all of the R code to perform the
analysis and make all of the figures. Given the output from a bcbio-nextgen run
and the Rmarkdown report, a researcher can regenerate and tweak the analysis and
figures for a project.
1.2
Discussion
This chapter described an open source, community curated analysis for RNA-seq
data that is being used by researchers in academia, core facilities, startups and pharmaceutical companies around the world. It builds on top of and extends previously
implemented infrastructure work that supports the automated deployment of all necessary software and data necessary to run an analysis. It is highly scalable, capable
of running thousands of samples simultaneously across a wide range of computing
environments including local machines, high performance clusters and on Amazon
and other cloud computing platforms. It provides a set of validation tools to justify
the implementation choices and to allow for future new methods to be compared
to existing methods in an unbiased manner. It implements features not found in
preexisting commercial and open source solutions, including filtering of isoform calls
by using the consensus of multiple tools, calling and classifying variants, and calling
existing RNA-editing events from RNA-seq data. It can be configured to work in
concert with a laboratory information management system (LIMS) to automatically
kick off analyses and can be configured to automatically upload results to Amazon S3
or to a Galaxy instance for dissemintation to collaborators. A report designed as a
47
1.3. FUTURE DEVELOPMENT
48
summary and a starting point for more in depth analysis is automatically generated
at the end of the process.
Importantly, this work is open source and under active development by a vibrant,
thorough community of next generation sequencing experts. As part of an umbrella
project, features that get incorporated into the infrastructure by the community,
such as adding the ability to use Docker images, are able to be used by all analysis
implementations. This ensures that as progress moves forward the analyses will keep
up with the current science.
1.3
Future development
There are many areas of active development, both in terms of the infrastructure
driving bcbio-nextgen and bcbio.rnaseq and improving the accuracy and speed
of the implemented analyses. An important feature that is missing from the RNAseq pipeline is support for finding and quantitating both known and novel very small
exons called microexons. Recent work has found microexons to be misregulated in
the brains of autistic patients and microexons to be preferentially expressed in brains
in general[79]. With the kick off of the BRAIN initiative soon to start, there should
be enormous interest in assaying transcription in the brain and microexons are likely
to be an important feature of the transcriptome to assay.
In addition, single-cell RNA-seq has become an exciting new technique and analyzing the data from single-cell experiments is challenging[8 1]. There is a large increase in technical noise caused by incomplete sampling of the transcriptome during
library preparation from a single cell[68]. The most sensitive protocols only capture
40% of the RNA molecules in each cell and so there are many artificial zero counts
due to the incomplete capture[68].
bcbio-nextgen and bcbio.rnaseq have been
1.3.
FUTURE DEVELOPMENT
49
used to analyze single-cell data but at present the implementation is not robus and
validation data is not available. As this type of data becomes more common, there
will be a need to develop an option to handle single-cell RNA-seq data.
We are in the process of swapping in much faster tools for transcriptome reconstruction[135]
and quantification[23][130]. These tools will allow for these tasks to be done on the
order of minutes instead of hours. The current work has highlighted the importance
of having validated, reproducible analyses. We have discovered bugs that escaped
reviewers in some of these programs that severely affected their output (see [134]).
Finally, the bcbio-nextgen community has been watching and participating in
the formation of the common workflow language (CWL). The CWL is an in-progress
specification for a language of describing data-intensive workflows in a language
agnostic way. The common workflow language is a high level language for describing
steps in a workflow by specifying the expected inputs and outputs and the tools
to run and linking them together to form an analysis pipeline [162].
This allows
the informatics workflow to be abstracted away from the compute and storage and
allows workflows to be shared across different infrastructure.
We plan to replace
the IPython.parallel method of parallelizing a bcbio-nextgen workflow with the
CWL when it is complete.
Chapter 2
Transcriptome defects in a
mouse model of tuberous
sclerosis
2.1
2.1.1
Background
Tuberous sclerosis
Tuberous sclerosis is an autosomal dominant genetic disorder characterized by growth
of benign tumors of the skin (angiofibromas), lung (lymphangiomyomatosis), heart
(cardiac rhabdomyomas), kidney (renal angiomyolipomas) and brain (cerebral cortical deformations called cortical tubers).
Tuberous sclerosis affected individuals
have a normal lifespan but have debilitating neurological symptoms including autism,
epilepsy and delays in intellectual development [84] [42]. Clinical signs of tuberous
2.1. BACKGROUND
sclerosis vary between patients with no one sign being diagnostic, but 80 % of patients present with cortical tubers. Tuberous sclerosis, Latin for 'hard swelling' is
named after this characteristic symptom of the disease.
Figure 2-1: CNS manifestations of tuberous sclerosis. A) an MRI showing several cortical
tubers with the apex of the tubers pointing towards the ventricle. This shape is due
to disruption in migration of cells from the ventricle to their proper cortical layer. B)
shows another neurological tumor that can develop, a subependymal giant cell astrocytoma
(SEGA). This is a non malignant tumor in the lateral ventricle. C) shows a magnified area
of a cortical tuber stained with an antibody against the translation marker ribosomal S6.
Arrows point to giant cells with overactive translation. Reproduced with permission [42],
Copyright Massachusetts Medical Society.
The cortical tubers are areas of the brain that develop abnormally due to disruption of outward migration of presumptive cortical cells from the ventricle. This
disruption of migration causes the characteristic triangular shape of the cortical tuber, with the apex pointing at the ventricle (Figure 2-1). The tuber itself is made
up of cells called giant cells which have large, swollen cell bodies and stain strongly
for ribosomal S6 protein, indicating overactive protein translation. The giant cells
can also stain for neuronal or glial markers, and occasionally for both markers, indicating the giant cells are undifferentiated or partially differentiated neurons or glia
[117].
In addition to the presence of giant cells, the area within and around the
51
2.1.
BACKGROUND
tuber shows disruption of normal cortical lamination, a phenomenon called cortical
dysplasia. This abnormality occurs in other diseases such as focal cortical dysplasia
and is most often associated with seizures. Intracranial EEG studies of children with
tuberous sclerosis confirm the link between the tubers and seizures. Subdural electrodes placed over tuberous and non-tuberous cortical tissue indicate that the tuber
itself is often, but not always, the locus of seizure generation Resection of tubers and
surrounding dysplastic cortex often halts intractable seizures, further supporting the
connection between tubers and seizure initiation. [118]. However, despite the strong
association between cortical tubers and seizures, some epileptic tuberous sclerosis patients do not have any discernable cortical tubers or cortical dysplasia [185]. Mouse
models of tuberous sclerosis also reflect this heterogeneity in the seizure phenotype
as the presence of both seizures and tubers can vary depending on the nature of the
knockout. Seizures without tubers or gross cortical abnormality suggest that changes
at the molecular level may also be contributing to the seizure phenoptype.
Tuberous sclerosis is a genetically simple syndrome. In 85 % of cases of tuberous
sclerosis, the root cause can be linked to mutations in one of two genes, TSC1 and
TSC2, that encode the proteins hamartin and tuberin, respectively. These two proteins bind together to form the tuberous sclerosis complex (TSC) that governs the
rate of translation in the cell via inhibition of the mammalian target of rapamycin
(mTor) pathway. Tuberous sclerosis is a monogenetic disorder because loss of function mutations in TSC1 or TSC2 cause the mTor pathway to be overactive and
allows protein translation to run out of control. Specifically, the TSC governs the
rate of protein translation through inhibition of Rheb-GTP activity and loss of TSC
function causes Rheb-GTP to be overactive.[99] Activated Rheb-GTP triggers the
mTORC1 kinase branch of the mTor pathway, which acts as a positive regulator
of protein synthesis by phosphorylating p70S6 kinase and 4E-binding protein 1 (4E-
52
2.1. BACKGROUND
BP1), both are involved in governing the rate of translation in the cell (Figure 2-2
on the next page). Phosphorylated p70S6 kinase in turn phosophorylates ribosomal
protein S6 kinase (S6K1), an important step in synthesis of the translation machinery
[66] while phosphorylation of 4E-BP1 causes it to dissociate from eukaryotic translation initiation factor 4E (eIF4E), allowing protein synthesis to occur[88]. Thus these
two phosphorylation events are essential to activation of the mTor pathway. They
cause the synthesis of a key component of the scaffolding of translation and also
releases a major translation factor from inhibition. Thus the pathway consequently
has a profound effect on the fate of the cell.
Increased protein production causes benign tumor-like growths called hamartomas in tuberous sclerosis patients through causing cells to either fail to enter
GO and remain preliforating or leave GO and re-enter the cell cycle[159]. This results in the formation of the hamartomas such as angiofibromas in the tuberous
sclerosis patients. Inhibiting mTor with rapamycin slows the rate of growth of these
hamartomas in tuberous sclerosis patients[75]. Deactivation of the tuberous sclerosis
complex not only affects the proliferation but also the differentiation of cells. This interference with differentiation is the cause of the poorly differentiated giant cells seen
in the cortical tubers and treatment with rapamycin can cause neurons to properly
differentiate after Tsc2 inhibition[158].
In addition to disrupting the migration and differentiation of neurons and glia,
overactivation of the mTor pathway has more subtle effects on many aspects neural
circuit formation ranging from axonal arborization to spine formation and synaptic
potentiation. These molecular and circuit effects may be the locus of the more subtle
neurological symptoms of tuberous sclerosis such as autism. Overexpression of either
Tscl or Tsc2 in cultured hippocampal neurons greatly reduces the number of neurons sprouting axons while knocking down either Tscl or Tsc2 has the opposite effect,
53
2.1.
54
54
BACKGROUND
2.1. BACKGROUND
I
I1
K I
protein synthesis
LLTP/LTD
cell survival/proliferation
actin rearrangement
spine formation
differentiation
Figure 2-2: The mTor pathway is affected in many disorders with autism as a phenotype.
Genes colored blue are genes known to be associated with autism. TSC1 and TSC2 are
disrupted in tuberous sclerosis, NFl in neurofibromatosis, FMRP in fragile X syndrome
and mutations in PTEN, receptor tyrosine kinases and NMDA receptors are all known to
be disrupted in some autism spectrum disorders.
2.1.
BACKGROUND
55
causing, in culture, an increase in the number of neurons with multiple axons[39].
Knocking out Tscl also disrupts myelination of neurons[114]. These results demonstrate that disruption of the activity of the tuberous sclerosis complex affects the
structure and functional capabilities of neural circuits. Supporting this notion, other
forms of local circuit disruption are found when the tuberous sclerosis complex is
not fully functional. Tsc2+/- mice have an aberrant retinotopic map which is formed
by a process of axon guidance to the correct termination zone via interaction of the
ephrins and the Eph receptors (see Chapter 3 on page 87) of this work for a brief
review)[127], indicating that proper regulation of the mTor pathway is necessary for
both the outgrowth of the axon and guidance of the axon to its proper termination zone.
Overactivation of the mTor pathway can also affect the fine tuning of
the synaptic contacts. It has long been known that blocking the NMDA receptor
with low levels of ketamine can induce new synapse formation in a protein synthesis
dependent manner. Blocking the mTor pathway with rapamycin blocks this effect
of ketamine on new synapse formation [96].
More subtly, mTor activation is im-
portant in protein synthesis dependent long long-term potentiation and LTD. Mice
heterozygous in loss of function mutations of Tscl or Tsc2 have a reduced threshold
for protein synthesis dependent long long-term potentiation[52] due to overactivation
of the mTor pathway. The decreased threshold for long long-term potentiation in
tuberous sclerosis heterozygote knockout mice is likely the cause of the deficits in
learning which can be reversed with rapamycin treatment. Knocking out Tscl can
also affect signaling in the neuropil through overactivate translation in glia. A Tscl
knockout specific to glia has excessive glutamate signaling which can be reversed
by partially blocking the NMDA receptor[192]. Disrupting Tscl function in neurons
causes a weakening of inhibition in the hippocampus, leading to overexcitability[12].
These results show that disruption of the careful balance of translation in the cell
2.1. BACKGROUND
56
has wide ranging consequences for the architecture of the brain and can affect many
phases of wiring of the neuropil, from the initial migration of the cells to their proper
positions, to controllng and guiding the axons via chemotactic signals. The defects
involve both the initial wiring of the brain, as with the retinotopic map disruption,
and the fine tuning of established circuits, as with the learning deficits. These multiple levels of brain connectivity disruption from overactive protein synthesis have
devastating neurological consequences for affected individuals.
There is no cure for tuberous sclerosis and treatments are focused on management
of the most severe symptoms. A promising avenue is chronic treatment with low
levels of rapamycin, the inhibitor of the mTor pathway that the mTor (mammalian
target of rapamycin) pathway is named after. Treatment with rapamycin can slow the
growth of tuberous sclerosis tumors, [53], help control epilepsy[25] reverse autism-like
symptoms and, in a Tsc2+/- mouse model of tuberous sclerosis, reverse short term
learning deficits[151] [52]. The success of rapamycin treatment on the symptoms of
tuberous sclerosis has caused an interest in investigating other treatments targeting
both the mTor pathway and pathways affected downstream of mTor[150]. To enable
these studies, many mouse models of tuberous sclerosis have been developed which
display subsets of the anatomical and neurological symptoms of tuberous sclerosis,
depending on the nature of the knockout.
Tscl and Tsc2 null mice are embryonic lethal, with renal tumors and failure
to close the neural tube[91], so much of the work on rodent models of tuberous
sclerosis focuses on less severe conditional knockouts and heterozygotes.
Condi-
tional knockouts of Tscl and Tsc2 have been made in both glia and post-mitotic
neurons. Glia-specific Tscl/2 knockouts were created by crossing floxed Tscl/Tsc2
alleles to mice expressing Cre recombinase under control of the human glial fibrillary acidic protein promoter (GFAP-Cre) to generate the glia-specific Tsc knockout
2.1.
57
BACKGROUND
mice (GFAP-Tsc1/2+)[177][193].
Neuron-specific Tscl/2 knockouts were created
by crossing floxed Tscl/Tsc2 mice to mice expressing Cre recombinase driven by
the rat synapsin I promoter to create Syni-Cre mice to generate neuron specific
knockouts (Syn1-Tsc1/2+/-)[114]. Interestingly, both the GFAP-Tscl~/- and GFAPTsc2~/-knockout mice both have seizures and disruption of the laminar structure of
the hippocampus, despite a lack of tuber formation[193]. The Syn1-Tsc1-/- mouse
has seizures, autism-like symptoms but no tubers or disruption of cortical lamination
(Figure 2-3 on the next page). The severity of the seizure and anatomical phenotype is temporally related to when Tscl function is lost in neurons, as a Emx1-Tsc1-/~
conditional knockout, that removes Tscl expression in neural progenitor cells, has
seizures and severe deficits in cortical lamination[32]. Since proper mTor activity is
necessary for neurons to differentiate and leave the cell cycle, overactive mTor prevents neurons from properly differentiating and migrating, resulting in the creation of
the poorly differentiated giant cells, cortical tubers and lamination disruption. The
Synl-Tscl-/- mouse loses Tscl after cells have differentiated and migrated to their
final location, so the cortical tubers and cortical dyplasia never form.
The above is a summary of the major neurological effects caused by overactivation
of the mTor pathway in tuberous sclerosis. These effects operate simultaneously on
several different levels of neuronal organization. At the highest level, aberrant mTor
activity affects differentiation of neurons and glia, resulting in cortical tubers, major
disruption of the cortex and seizures. At this level the neurons that make up the
neuropil are not properly formed or not properly localized. Knockouts where the
activity of the tuberous sclerosis complex is disrupted either later on in development
or in a subset of already differentiated cells have phenotypes that are a subset of
the tuberous sclerosis symptoms. In these milder models the more subtle aberrant
wiring of the neuropil is revealed, with disruptions both in chemotactic and activity
58
BACKGROUND
2.1.
58
-
2.1. BACKGROUND
.. ,
,
.k
4
-~
4
~
*
4
q
~E..
~
.
4
Figure 2-3: The laminar structure of the cortex in Syn1-Tscl~/- mice is undisturbed.
cells and
Syn1-Tscl/- mice, a neuron-specific knockout of Tscl, has seizure activity, giant
of
structure
laminar
the
dysmorphic neurons and glia, but no tubers and no disruption of
B2)
and
the cortex. B1) NeuN staining shows a normal cortical structure in the controls
Syn1-Tsc1-/- mice. cresyl violet staining also shows no difference between Cl) controls
and C2) Syn1-Tsc1-/- mice. Used with permission from [185], copyright John Wiley and
Sons.
2.1.
BACKGROUND
59
dependent wiring. These two features of tuberous sclerosis, that it is a monogenic
disorder causing autism and that mice can be generated which show only a subset
of the phenotypes make it a powerful model system for the study of autism.
2.1.2
Tuberous sclerosis is a tractible autism model
Autism spectrum disorder is a complex multigenic disorder with evidence of hundreds of succeptibility loci that, with some exceptions, only confer a small amount of
risk. A recent whole genome study of quartet families where both parents and two
ASD affected siblings were sequenced was a dramatic demonstration of the genetic
complexity:
siblings from the same family, both with autism spectrum disorders
were found not to share the same putatively causitive mutations in over 70 % of the
cases. [191]. Most diagnosed autism disorders are genetically complex with no single
causative mutation, however there are monogenic diseases which feature autism as
one of the predominant symptoms. These monogenic disorders include neurofibromatosis, mutations in Pten, Fragile-X disorder and tuberous sclerosis. Monogenic
disorders causing autism are valuable as systems to study and treat the symptoms
of autism spectrum disorders as knockout animal models can be generated that have
an autism phenotype[10].
The efficacy of autism spectrum disorder interventions
can then be assayed using a social behavioral test for autism spectrum disorder in
mice[157].
Many of the monogenic disorders with a penetrant autism phenotype involve
disruption of genes regulating the mTor pathway (see Figure
2-2 on page 54).
Over 50 % of patients with tuberous sclerosis present with symptoms of autism
spectrum disorder (ASD). Mouse models disrupting Tscl or Tsc2 function recapitulate the social deficits and can be rescued with low levels of rapamycin[166]. Mu-
2.1. BACKGROUND
60
tations in NFl, the gene coding for neurofibromin, cause neurofibromatosis type
1, with 15-30% of affected individuals presenting with autism spectrum disorder
symptoms[63][140]. Neurofibromin inhibits Ras-GTP and Ras-GTP activates Erkl/2
which inhibits Tscl/2, leading to overactive mTor. A mouse model knocking out Nfl
has an autism phenotype that is rescued by blocking Pakl[119], a kinase activated by
the mTor pathway[69]. Pten mutations cause autism in 10 % of affected patients[61]
and again act through the mTor pathway. Pten loss of function leads to overactivation of AKT which blocks the activity of the tuberous sclerosis complex. A mouse
model deleting Pten has autism-like behavioral deficits[102], and a neuron specific
knockout of Pten has cortical dysplasia and seizures which, like tuberous sclerosis,
can be rescued by rapamycin treatment [126].
Finally, Fragile X syndrome is an-
other disorder presenting with autism like symptoms through activation of the mTor
pathway, via disruption of the fragile X mental retardation protein (FMRP). FMRP
normally inhibits P13 kinase, and release of P13 kinase from inhibition causes the
TSC to be inhibited[156]. All of these autism disorders increase activity of the mTor
pathway through indirect inhibition of the activity of the tuberous sclerosis complex
(Figure 2-2 on page 54).
The tuberous sclerosis complex is the common gateway through which these monogenic autism disorders exert their effects. The difficulty with using tuberous sclerosis
as a model for studying autism is that tuberous sclerosis, while genetically simple,
is phenotypically complex. Patients present with a wide range of severe symptoms
and teasing apart what is involved in each phenotype is difficult. This complexity
is ameliorated, however, by the existence of mouse models that have only a subset
of the tuberous sclerosis symptoms. In particular, Synl-Tscl-/- mice lack the gross
anatomical defects of other tuberous sclerosis models, but retain the seizure and
autism phenotypes. This mouse represents the opposite of the autism genome-wide
2.1. BACKGROUND
association studies analyses are restricted to patients with non-syndromic autism
with as few comorbid phenotypes as possible but one where the underlying genetic
cause is unknown and complex. Studying Synl-Tscl-/- mice trades genetic complexity for phenotypic complexity, but the phenotupic complexity is reduced by studying
the neuronal knockout of Tscl which has only the seizure and the autism spectrum
disorder phenotypes.
2.1.3
The second order effects of mTor activation are important
Despite these autism associated disorders all resulting in overactivation of the mTor
pathway, the downstream effects of mTor activation are not the same. This can
be seen at multiple levels. At the phenotypic level, tuberous sclerosis disorder, Cowden's syndrome, where PTEN is lost, and neurofibromatosis all present with seizures,
hamartomas or tumors and autism-like symptoms. Fragile X syndrome, on the other
hand, presents with seizures and autism but the hamartomas are absent. The penetration of each endophenotype differs amongst these disorders with varying rates
of autism, tumor and seizures in the population. This indicates that either downstream effects or non-mTor related effects of each disorder are driving phenotypic
differences. At a molecular level, these syndromes can have vastly different effects despite the common overactivation of the mTor pathway. Multiple studies have shown
the importance of dissecting the second order effects specific to each syndrome.
The effect of fragile X syndrome and tuberous sclerosis on the mGluR5 receptor
is a salient example. Fmrl-/y mice, a mouse model of the fragile X syndrome, have
enhanced mGluR5 dependent LTD which is triggered by enhanced protein synthesis
downstream of mGluR5 activation. Tsc2+/- mice also have excessive mTor activation
61
2.1. BACKGROUND
62
but have decreased protein synthesis dependent mGluR5 LTD[10]. Crossing Tsc2+/and Fmrl-/y mice to each other results in normal mGluR5 activity indicating that
these two mutations have opposite effects on mGluR5 activation and can be balanced
out, despite both resulting in overactivation of the mTor pathway.
Both of these
knockouts have phenotypes which can be rescued with rapamycin treatment, but this
result shows that there are secondary effects of the mTor pathway activation that
are specific to a particular genotype. In the case of Fmrl-/y and Tsc2+/, therapies
targeting mGluR5 in one disorder would be contraindicated for the other disorder,
despite both affecting the same pathway. A more nuanced dissection of the specific
molecular effects of overactivating the mTor pathway is important if more focused,
disease specific therapies are to be developed.
Most work assaying gene expression differences in autism has focused on large cohorts of non-syndromic autism disorders. Those studies have found differing amounts
of evidence for large scale gene expression changes in the brain across non-syndromic
autism disorders. A study of moderate size comparing gene expression in postmortem
control and autistic human cortices found only MAL and C11ORF30 differentially
expressed, both at very moderate fold changes. Pathway analysis in the same study
found a gene co-expression module upregulated for immune-related genes[70]. Another study that examined both human cerebellum and cortex found there was scant
evidence for gene expression differences in cerebellum but hundreds of differentially
expressed genes in the frontal and temporal cortex of autism affected individuals[180].
Limited work has been done exploring the effect of mTor-linked syndromic autism
disorders on transcription in the brain. A study examining the transcriptome of
the cerebellum of Tsc2+/- and Fmrl-/y mice and found no gene-level significant differences in any comparisons, including Tsc2+/- and Fmrl-/y compared to control
mice[92]. This study focused on the cerebellum which has mixed evidence for gene
2.2.
METHODS
expression differences in autism spectrum disorder brain[180]. Despite the diverse
results, characterizing the transcriptome of specific syndromic autism spectrum disorders is important. Just as cancer is now known to not be a single disorder but to
have specific subtypes that have very different molecular pathways disrupted, it is
likely that a complex, subtle neurological disorder such as autism also can be caused
by multiple different pathways. The differing regulation of mGluR5 in fragile X syndrome and tuberous sclerosis demonstrates the importance of characterizing these
disorders at the molecular level.
For this study we used RNA-sequencing to assay gene expression changes in the
frontal cortex of Syn1-Tsc1-/- mice that have Tscl knocked out only in neurons[114]
via use of the rat synapsin I promoter, a membrane-associated protein expressed only
in neurons. We examined differential gene expression, differential splicing and differential RNA-editing between wild type and Synl-Tscl-/- mice and found hundreds
of gene and splicing differences. We found a set of genes known to be involved in
autism and epilepsy which are differentially transcribed in these mice. Following up
on one of these hits, we show evidence that overexpression of the serotonin receptor
Htr2c causes hyperexcitability and synchronized calcium influx into the neuron.
2.2
2.2.1
Methods
Cortex collection
Synl-Tscl-/- mice were prepared by crossing Syn1-Tsc1loxP+/--Syn1Cre+/- mice and
Syn1-Tsc1loxP+/--Syn1Cre-/- mice of the opposite sex. Four week old Synl-Tscf-/~
and unaffected control littermates were anesthetized with isoflurane and decapitated.
The frontal cortex was dissected, rinsed with ice-cold PBS and stored in RNALater
63
2.2.
METHODS
at -20 C after being cut into 3-4 mm pieces to ensure penetration of the RNALater.
Synl-Tscl-/- mice had phenotypes as previously described[114], with tremors, a hindlimb grasping reflex when suspended by their tail, seizures and a much lower weight
than control littermates. In addition, Synl-Tscl-/- mice, when spun gently by their
tails often had a a very large seizure resulting in death. Genotypes were confirmed
with PCR using a standard genotyping protocol and a set of previously described
primers[114].
All experiments were carried out with the approval of the Committee on Animal
Care at the Massachusetts Institute of Technology.
2.2.2
Library preparation and sequencing
Cortices were placed in 1 mL Qiazol and homogenized and total RNA was extracted
following the manufacturers protocol. Contaminating genomic DNA was removed
from the total RNA by treatment with ten units of DNAse 1 (Roche) following
the manufacturers protocol. Total RNA was further cleaned up using the RNeasy@
MinElute@ cleanup kit (Qiagen #74204) to remove contaminants left over from the
extraction process following the manufacturer's protocol with the exception of rinsing
the pellet with ethanol three times. Total RNA was eluted in 30 p1 of RNAse free
water. The purity of the samples was analyzed using a NanoDrop spectrophotometer
(Thermo Scientific) and samples with 260/280 ratios less than 1.8 or 260/230 ratios
less than 2.0 were subjected to a second round of cleanup. After cleanup the samples
were run on an Agilent 2100 Bioanalyzer and samples with a RNA integrity number
of less than 9, which indicates degradated RNA, were discarded. Purified total RNA
samples were stored at -80 C until library creation.
Libraries were created for paired-end sequencing on the Illumina HISeq using
64
2.2. METHODS
65
the Illumina TruSeq v2 kit, following the manufacturer's protocol. Libraries were
created in two batches, with equal control and Syn1-Tsc1-/- samples in each batch.
Equal samples from control and Synl-Tscl-/- litters were included in each batch. For
sequencing, in one batch, two control and two Synl-Tscl-/- samples were multiplexed
on a single lane. In the other batch, a single control and Synl-Tscl-/- sample were run
in separate lanes. Samples were sequenced at the BioMicroCenter at MIT, sequencing
40 basepairs from each end of the reads.
2.2.3
Informatics analysis
Alignment and differential expression
The RNA-sequencing reads were processed using the RNA-seq pipeline implemented
in version 0.8.3 of the bcbio-nextgen analysis project. Briefly, poor quality bases
with PHRED scores less than five[105], contaminant adapter sequences and polyA
tails were trimmed from the ends of reads with cutadapt[109] version 1.2.1, discarding
reads shorter than twenty bases. A STAR[48 index was created from a combination
of the Mus musculus version 10 (mm10) build of the mouse genome and the Ensembl release 75 gene annotation. Trimmed reads were aligned to the STAR index,
discarding reads with ten or more multiple matches to the genome.
Quality met-
rics including mapping percentage, rRNA contamination, average coverage across
the length of the genes, read quality, adapter contamination and others were calculated using a combination of FastQC, RNA-SeQC[45] and custom functions from
bcbio-nextgen and bcbio.rnaseq. Chapter 1 on page 11 of this work has more specific
implementation details.
The features of the transcriptome were quantitated at the gene, isoform and
exon level by three separate methods.
Reads mapping to genes were counted us-
2.2.
METHODS
66
ing featureCounts[97] version 1.4.4, excluding reads mapping multiple times to the
genome and reads that could not be uniquely assigned to a gene. Reads mapping
to isoforms were counted using eXpress[147].
Reads aligning to individual exons,
broken up into unique features, were counted using DEXSeq[7] version 1.12.1.
Differential expression at the level of the gene was called using DESeq2 version
1.6.3, filtering for genes using a BH corrected[15] false discovery rate (FDR) cutoff
of 0.1. Splicing differential expression was called at the isoform level with EBSeq
version 1.6.0, using a matrix of normalized effective counts for each isoform from
eXpress and at the exon level using DEXseq. To control the false positive rate when
calling differential isoforms, a transcript was called differentially spliced only if both
EBseq and DEXseq called a differential splice in the same transcript with a FDR
cutoff of 0.1.
A background set of 15,807 expressed genes in the cortex was constructed by
filtering the normalized count data for genes with a mean count of at least ten. A set
of transcriptionally regulated genes was made by compiling all genes differentially expressed with all genes with a splicing event, in total 407. The set of transcriptionally
regulated genes were tested for overrepresentation in pathways using the expressed
set of genes as a background using WebGestalt[89] with a FDR cutoff of 0.1.
RNA-editing
Using bcbio-nextgen version 0.8.3, STAR aligned reads were preprocessed by splitting alignments with deletions into two separate alignments to prevent the variant caller from calling spurious deletions at exon-exon junctions[78].
The geno-
types of the preprocessed reads of individual samples were called with the GATK
HaplotypeCaller[44] which was configured to output genomic variant calling format
67
2.2. METHODS
(gVCF) intermediate files for joint calling at all positions with a variant in any
sample[178]. These files contain both variants and probability information of reference calls. These intermediate files were then jointly called with GenotypeGVFs to
produce a final set of squared off variants for all of the samples. Known mm9 coordinates of RNA editing sites from the DARNED and RADAR databases were lifted
over to mm10 coordinates and combined into a set of 17,831 curated RNA editing
sites. A-÷G and T-C variant calls from HaplotypeCaller were intersected with the
curated editing sites to produce the final set of edited sites. The functional effect of
the RNA edits was annotated with Variant Effects Predictor[190].
There is a lower level of editing that we can reliably detect. Assuming editing
follows a binomial distribution, we need to have at least n = 4p6 -p) observatons to
detect at least one edited read 95% of the time for a fraction of edited transcripts p.
We set the edit fraction to 0.1 and filtered out any editing events with less than 36
total observations in either the control or Syn1-Tsc1~/
samples to keep a set where
we can reliably detect editing, if it is occuring, in both conditions. Editing events
were flagged as differentially edited via the binomial test with a cutoff of FDR < 0.1
and a editing percentage difference > 25 %.
2.2.4
Calcium Imaging
At embryonic day 15.5, fetal brain cortices of Synl-Tsc1-/- mice were dissected and
collected. Concurrently, the rest of the brain tissue were used for genotyping after
DNA was extracted using a commercial kit (Sigma Aldrich #XNAT2).
Fetal cor-
tices were digested with a solution containing papain and Dase for 25 min. Cells
were dissociated using pipets. Cells were plated to a coverslip coated with laminin
and poly-D-lysine. A DNA construct encoding GCaMP3 (a gift from Loren Looger,
2.3.
RESULTS
Addgene plasmid #22692)
68
was transfected at day 6-8 in vitro using Lipofectamine
2000 (Invitrogen) according to the manufacturer's protocol. Neurons were imaged
at DIV24-30 using a 40x objective on a Nikon Eclipse E600FN confocal microscope.
Tyrode solution without MgCl 2 , was used, which consisted of (in mM) NaCl, 145;
KC1, 3.0; HEPES, 10; Glucose, 10; Glycine, 0.005; CaCl 2 , 2.6; 300 mOsmol; adjusted
to pH 7.4. Serial images were taken at 512 x 512 pixels using a water immersion
lens with a large pinhole. In a pilot experiment, various intervals (0.125-1 see per
frame) were tested and patterns of GCaMP3 signal intensity changes were comparable. Consequently, the interval at 1 see per frame was chosen. For pharmacological
treatments, neurons were treated with 0.1pmol SB242084 (Tocris #2901)
for 1-2
hours after baseline images were captured. In the second experiment, neurons were
treated with 5 pmol Ro25-6981 (Tocris #1594) for 1 hour, then 0.1 pmol SB242084,
and 50 pmol D-APV (Sigma-Aldrich #A8054) were sequentially added to the imaging solution at the interval of one hour between imaging. Ro25-6981 blocks the NR2B
subunit of the NMDA receptor, SB242084 blocks the HTR2C receptor and D-APV
blocks the NMDA receptor.
2.3
2.3.1
Results
Differential expression
We identified 251 genes at FDR < 0.1 as differentially expressed in the cortex of
Syn1-Tscl-/- mice compared to wild type littermates. (Table 2.1 on page 75). We
also found 254 differential splicing events in between the two conditions.
EBSeq
called 2895 isoforms differentially expressed (FDR < 0.1, representing 2436 genes,
whereas DEXseq called 549 exons differentially expressed at FDR < 0.1, representing
2.3.
RESULTS
489 genes.
69
Taking consensus calls between EBSeq and DEXseq left us with 254
differential splicing calls in 156 genes.
CD)
0-
_
0
00-q.%
00
0
LO
0
0
00
%
3L,
3L..
0
to
Z)
0)
0 7-0
0
000
00.
0
0
0
0)
woo
4& -W
Oevij-,
we
0
0 0 .00% 0 0 0
00
:
0
, 0
U)
C0
0
0.0 0 0
0
0 000
*0
0
0 0
0O
0.1
10
1000
10000
mean expression
Figure 2-4: MA-plot of gene expression in the cortex of wild type vs. Syn1-Tsc1-/- littermates. 251 genes are differentially expressed at a false discovery rate of 0.10 (colored red).
Genes expressed more highly in Tsc-/- mice have a positive log2 fold change.
2.3. RESULTS
70
symbol
log2 fold change
FDR
symbol
log2 fold change
FDR
Collal
0.89
0.00
Vwf
-0.56
0.09
Rec8
-1.09
0.00
Il17ra
0.58
0.10
Rab3b
-0.69
0.00
Ddrl
0.56
0.05
Fosb
0.70
0.04
Calr
0.46
0.09
Grm3
0.67
0.01
Hif3a
-1.38
0.00
Sst
-0.78
0.00
Nes
0.85
0.00
Wispl
0.71
0.07
Ephb3
0.59
0.04
Cbfa2t3
1.08
0.00
Adamts4
0.77
0.01
Itih3
-0.62
0.08
Cnp
0.94
0.00
Homeri
0.79
0.00
Dio2
0.55
0.05
Hnrnphl
0.47
0.10
Bcasl
0.61
0.02
Ferls
1.27
0.00
Phf2lb
0.61
0.07
Scubel
0.63
0.02
Pltp
0.70
0.01
Erbb3
0.91
0.00
Ksrl
-0.84
0.00
Slc13a3
0.64
0.09
Kdm6b
0.52
0.10
Dusp14
0.72
0.00
P4hal
0.75
0.00
Den
0.67
0.09
Dusp6
0.61
0.01
Epb4.112
0.50
0.08
Cfap54
-1.07
0.00
Mybpcl
-1.78
0.00
Unc5b
0.83
0.00
Ddit4
-0.70
0.01
Gamt
0.84
0.01
Cobl
0.49
0.08
Pgam2
-0.92
0.00
Trib2
0.69
0.01
Cacnalg
0.69
0.00
Serpina3n
-1.15
0.00
Gaint16
-0.77
0.00
Asb2
-0.84
0.01
Akrlcl8
-1.16
0.00
2.3. RESULTS
71
symbol
log2 fold change
FDR
symbol
Net1
-0.61
0.04
Tppp
0.48
0.07
PeskI
0.66
0.02
Crhbp
-0.63
0.09
Galnt15
-1.00
0.00
Dmtn
0.47
0.09
Dct
-0.67
0.04
Kcnvl
0.56
0.03
Myh9
0.60
0.02
Emp2
0.76
0.01
Arc
0.80
0.02
Tfrc
-1.15
0.00
Nr4al
0.89
0.00
Synj2
0.66
0.00
Xdh
-0.86
0.01
Col11a2
0.97
0.00
Ttc39c
-0.66
0.08
Nr3cl
0.50
0.07
Slc22a6
0.85
0.01
Tnfrsf25
1.23
0.00
Drapi
-0.52
0.08
Gfral
0.64
0.05
Alas2
0.82
0.01
Gil
-0.76
0.04
Cpebl
-0.65
0.01
Pdia4
0.73
0.01
Col3al
0.99
0.00
Cdk18
0.63
0.08
Tmem63a
0.67
0.05
Mgst3
-0.51
0.08
Myoc
1.95
0.00
Gad2
-0.54
0.04
Lcn2
-1.34
0.00
Ermn
0.98
0.00
Gsn
0.72
0.01
Sord
-0.67
0.07
Lamp5
0.67
0.00
Chgb
-0.51
0.04
Mal
0.58
0.03
Pdyn
-0.63
0.04
Cd93
0.76
0.02
Car2
0.56
0.05
Slc7a11
0.89
0.00
Serpinil
-0.51
0.05
Olfml3
1.23
0.00
Tspan2
0.55
0.06
T1rp53inp1
-0.64
0.10
Calbi
0.54
0.04
log2 fold change
FDR
2.3. RESULTS
symbol
72
log2 fold change
FDR
symbol
log2 fold change
FDR
Nr4a3
1.04
0.00
Podn
0.68
0.09
Hivep3
0.54
0.03
Atpafl
-0.61
0.06
Hspg2
0.76
0.01
Csmd2
0.61
0.01
Map3k6
-0.93
0.00
Sema3a
0.61
0.06
Cort
-0.99
0.00
Fos12
0.72
0.00
Ociad2
-0.83
0.00
Cenpa
-0.82
0.02
Tpst2
-0.57
0.09
Snx8
-0.64
0.07
Rph3a
0.50
0.09
Colla2
0.67
0.02
Gpnmb
-1.14
0.00
KIfM5
-0.61
0.07
Beati
-0.51
0.07
Dmpk
0.68
0.03
Sultial
-0.76
0.03
Dgat2
-0.55
0.05
Cox6a2
-0.75
0.05
Crym
-0.72
0.00
Plxnb3
1.01
0.00
Plp1
0.65
0.00
Cdhll
0.56
0.04
Pllp
0.54
0.08
Cx3cl
0.53
0.02
Mast3
0.49
0.04
Robo3
1.08
0.00
Mcam
0.95
0.00
Dock6
0.58
0.09
Myole
-0.61
0.06
Thsd4
-0.63
0.08
Coll2al
0.65
0.08
TribI
0.81
0.00
Ephbl
0.62
0.03
Ugt8a
0.85
0.00
Fa2h
0.95
0.00
Egr3
0.96
0.00
9430020K01Rik
0.49
0.09
ScnlOa
-0.91
0.00
Pla2g3
-0.99
0.00
Tdg
0.82
0.00
Tet3
0.50
0.08
Fam2l4a
-0.58
0.06
Eef2k
-0.52
0.09
2.3. RESULTS
73
symbol
log2 fold change
FDR
symbol
log2 fold change
FDR
Tanci
-0.67
0.02
Lars2
-0.48
0.08
Pdzrn3
0.56
0.05
Midn
0.66
0.00
Cerk
-0.49
0.09
Dnajb5
0.65
0.00
Tm6sf2
0.69
0.05
Lrrk2
0.57
0.06
P2ry12
0.86
0.00
Rnf39
0.71
0.06
Mag
0.94
0.00
Arhgap33
0.56
0.10
Mycn
0.83
0.01
Clic4
0.58
0.03
Klf10
0.95
0.00
Cldn11
0.80
0.00
Inf2
0.62
0.03
Vstm2l
0.68
0.00
Egr2
1.43
0.00
Ostml
-0.52
0.09
Sesni
-0.57
0.05
Pcp4ll
-0.60
0.03
Egri
1.14
0.00
Lparl
1.03
0.00
Gpr126
-0.81
0.02
Rnf122
0.71
0.07
Gpr37
0.78
0.00
Thbsl
0.91
0.00
Ankrd6
0.65
0.02
Dnm3
-0.85
0.00
Stac3
0.69
0.09
Bcor
0.58
0.07
Spns2
-0.58
0.04
Ltbp4
0.54
0.04
Chrm4
0.64
0.03
Sh3pxd2b
0.60
0.06
Eml2
-0.50
0.06
Dlk1
-1.02
0.00
Htr2c
-0.75
0.02
Fmod
0.63
0.08
Mbp
0.51
0.03
Pkp2
-0.71
0.07
Map3k9
-0.47
0.09
Gpr153
0.87
0.00
Nrep
0.60
0.02
Tubalc
0.89
0.00
Gjc2
0.93
0.00
Mdgal
0.67
0.01
2.3. RESULTS
symbol
74
log2 fold change
FDR
symbol
log2 fold change
FDR
SiprI
0.62
0.01
Cirbp
-0.60
0.02
Ifit2
-0.88
0.00
Lrrc75a
0.65
0.08
Mc4r
-0.83
0.01
Sstr2
0.71
0.02
Phldbl
0.85
0.00
Gpr137c
-0.57
0.08
Pdpl
0.55
0.02
Opalin
1.06
0.00
D8Ertd82e
1.07
0.00
Scg2
-0.81
0.00
Islr2
-0.67
0.00
Gpr17
0.49
0.10
Hbb-bs
1.22
0.00
Cx3crl
0.64
0.04
Hrhl
0.77
0.02
Cntn2
0.56
0.04
Mapkll
0.77
0.00
Spock3
-0.63
0.01
Nelli
0.65
0.01
Fmnll
0.66
0.00
Prr18
0.71
0.02
Trnpl
-0.74
0.00
Ier5
0.84
0.00
Myh4
-1.16
0.00
Cdh24
0.74
0.02
Plekhhl
0.73
0.01
Cxcl12
0.68
0.00
Ppplrlb
-0.64
0.01
mt-Nd6
-0.58
0.02
Cacng3
0.58
0.02
Zbtbl6
-0.59
0.03
D430041DO5Rik
0.55
0.02
Slc7al4
-0.69
0.00
Hba-al
1.05
0.00
SerpinhI
0.88
0.00
Gadi
-0.56
0.01
Egr4
0.81
0.00
Zhx2
0.65
0.09
Tmem88b
0.76
0.01
Klhl40
-1.01
0.00
Kcngl
0.98
0.00
Thbd
1.04
0.00
Fnbpl
0.50
0.09
Mog
0.77
0.00
Rab26
-0.63
0.07
Wipf3
-0.76
0.00
2.3.
RESULTS
symbol
75
log2 fold change
FDR
symbol
log2 fold change
FDR
Thncl
0.79
0.03
Plcxd2
0.50
0.06
Cecr6
0.56
0.04
Table 2.1: Differentially expressed genes in Syn1-Tsc1-/- mice. 251 genes are differentially
expressed with FDR < 0.1 in cortical neuron Synl-Tscl-/- knockout mice compared to
sibling controls. Negative log2 fold change are genes more highly expressed in the SyniTsc-/- mice compared to control mice.
2.3.2
Pathway analysis
The set of 407 differentially regulated genes is highly enriched for genes in the mTor
pathway compared to the background set of 15,807 expressed genes (hypergeometric
test, FDR < le-5), with an even distribution of genes regulated at the expression and
splicing level (Table 2.4 on page 78). KEGG[86], Gene Ontology (GO) molecular
function and and GO cellular part analyses show an enrichment in genes involving
axon guidance, signalling and extracellular matrix interaction (Tables
2.2 on the
next page and Table 2.3 on page 77).
2.3.3
Isoform differential expression
Since isoform differential expression is prone to generating false positive results, we
combined the calls from two independent methods into one consensus call set. Differential exon usage was called using DEXseq version 1.12.1 using counts generated from
featureCounts.
549 exons were differentially expressed at FDR < 0.1, representing
2.3.
RESULTS
76
KEGG pathway
ECM-receptor interaction
pvalue
FDR
0.0004 0.0048
genes
Colla2, Collal, Thbsl, Hspg2, Col3al, Collla2, Vwf
Protein digestion and absorption
0.0020
0.0120
Colla2, Coll2al, Collal, Col3al, Col11a2
Neuroactive ligand-receptor interaction
0.0033
0.0132
Grm3, Nr3cl, Htr2c, Sstr2, Chrm4, Hrhl, SIprI, Lparl, Mc4r
Axon guidance
0.0068
0.0204
Unc5b, Sema3a, Ephb3, Ephbi, Plxnb3, Cxcl12, Robo3
Amoebiasis
0.0477 0.0847
Colla2, Collal, Col3al, Coll1a2
Tight junction
0.0476 0.0847
Epb4.112, CldnI1, Myh4, Myh9, Rab3b
MAPK signaling pathway
0.0494 0.0847
Dusp6, Cacng3, Pla2g3, Mapkll, Map3k6, Nr4al, Dusp14, Cacnalg
Cytokine-cytokine receptor interaction
0.0620 0.0930
Calcium signaling pathway
0.04
0.08
Tnfrsf25, Cxcl12, Cx3cll, Cx3crl, Il17ra
Tnncl, Prkcg, Cacnalg, Htr2c, Atp2b2, Grini, Erbb3, Hrhl
Table 2.2: KEGG pathway analysis of differentially regulated genes. Several KEGG
pathways are enriched (FDR < 0.1) comparing the set of differentially expressed genes to
the background set of expressed genes. Pathways involved in cell motility, signal transduction in neurons, axon guidance and calcium signaling were differentially activated in
Synl-Tscl- mice compared to wildtype controls.
492 different genes. Differential isoform expression was called using EBSeq version
1.6.0 with effective isoform counts generated from eXpress version 1.5.1. EBSeq called
2895 isoforms differentially expressed representing 2436 genes. Consensus calls between EBSeq and DEXseq were constructed by only flagging transcripts where both
DEXseq and EBSeq predicted differential splicing as differentially expressed. This
process left us with 254 differential splicing calls in 167 different genes.
2.3.4
Disease association
The set of differentially expressed genes and the set of differentially spliced transcripts
were combined into a master set of genes undergoing transcriptional regulation. A
set of autism, epilepsy and intellectual disability associated genes were compiled
from DisGeNET[13], a database of gene-disease associations integrated from several
sources.
Table
2.5 on page 79 breaks down the genes that are transcriptionally
regulated in those diseases.
2.3.
77
77
RESULTS
2.3. RESULTS
GO term
pvalue
FDR
receptor activity
7.67e-06
0.0006
signaling receptor activity
8.35e-06
0.0006
transmembrane signaling receptor activity
2.65e-05
0.0007
signal transducer activity
3.20e-05
0.0007
molecular transducer activity
3.20e-05
0.0007
protein binding
2.15e-05
0.0007
G-protein coupled receptor activity
1.93e-05
0.0007
hormone binding
0.0001
0.0018
structural molecule activity
0.0002
0.0033
extracellular matrix structural constituent
0.0004
0.0053
extracellular space
2.49e-10
2.86e-08
extracellular region
2.83e-09
1.63e-07
extracellular region part
6.64e-09
2.55e-07
myelin sheath
2.72e-08
6.97e-07
axon
3.03e-08
6.97e-07
cell periphery
3.78e-08
7.25e-07
plasma membrane
1.32e-07
2.17e-06
cell surface
1.30e-06
1.87e-05
neuron projection
5.38e-06
6.19e-05
cell projection
5.36e-06
6.19e-05
Table 2.3: Gene Ontology (GO) term analysis of differentially regulated genes. Top ten GO
terms for molecular function and cellular component enriched in the set of differentially
expressed genes compared to the expressed background; in all there were 40 significant
cellular component terms and 26 significant molecular function terms. Terms for signal
transduction, interactions with the extracellular space, axon development and myelination
are overrepresented.
2.3.
RESULTS
78
78
2.3. RESULTS
gene
isoform
both
Thbd, Cxcl12, Gsn, Den,
Trf, Gm20425, Gm12117,
Cacnalg, Egri, Tfrc
Trp53inpl, Gfral, Slpri,
Dnml,
Ddit4,
Collal,
Gm3839,
Ksrl,
Dusp6,
Gapdh,
Sesni,
Egr2,
Mapkll,
Eef2k,
Fosb,
Gm11214,
Gm7293,
Iqsecl,
Plekha2,
Csnklg2,
Map3k6, Nr3cl, Serpinil,
3110039M2ORik,
Colla2, Nr4al
Ppargcla, Prkcg, Eif4gl,
Rblccl,
Foxgl,
Zfyve28,
Gnaol,
Exoci,
Gdil,
Eng,
MapklO, Hspa8, Arpc3,
Dnm2
Table 2.4: mTor signalling is differentially regulated in the Synl-Tsc-/- mouse. The mTor
pathway is overrepresented by genes with expression and splicing differences in the SynI-
Tsc-/- mouse, compared to the background set of expressed genes (hypergeometric test,
FDR < 0.0000135).
2.3. RESULTS
79
autism
Gpr37,
Htr2c,
Cuxi, Nrxn2
Zhx2,
intellectual disability
epilepsy
Fosb, Calr, Dio2, Gamt,
Fosb, Calr, Grm3, Sst,
Tppp,
Nes,
Nr3cl,
Car2 Sema3a,
Alas2,
Gpnmb,
Dmpk,
Sultial,
Zbtbl6,
Hba-al,
Gdil,
Inf2,
Dusp6,
Cacnalg,
Crhbp,
Gria3, Alas2,
Gamt,
Tppp, Pcskl,
Myh9,
Nr3cl
Gad2,
Lcn2,
Lamp2,
Foxgl
Pdyn,
Car2,
Eng,
Grini,
Fosl2,
Sultial,
Plpi,
Arhgef2, Wdr13, Dnm2,
Ugt8a,
Egri,
Thbsl
Dbnl,
Dlkl, Htr2c, Nrep, Sstr2,
Stxbpl,
Metap2,
Iqsec2, Mapk1O
Tbce,
Serpinil,
Cntn2, Cxcl12, Cacng3,
Gadi,
Lamp2
Gria3,
Csnklg2,
Foxgl,
Dapkl,
Eng,
Dnml,
Grini,
Tpm3,
Arhgef2,
Cuxi,
Gnaol,
Stxbpl,
Lphn3
MapklO
Table 2.5: Autism, intellectual disability and epilepsy genes that are differentially regulated
in the Syn1-Tsc1-/- mouse.
2.3.
RESULTS
80
non-coding (2851)
coding (54)
class
percent
class
percent
intron
38
missense
84
downstream
25
synonymous
15
upstream
12
in-frame deletion
1
non-coding
11
3'UTR
4
Table 2.6: RNA editing events found in Synl-Tsc-/ and WT mice. 3169 editing events
were identified in the WT and TSC samples, using known edit events from the DARNED
and RADAR databases for mm1O. VEP annotation revealed very few edit events with
functional effects, with most edit events in either introns or genomic regions flanking a
gene. Most edits in coding regions are missense mutations.
2.3.5
RNA editing
Distinguishing RNA-editing events from germline variants is a difficult problem,
prone to generating many false positives[11] so we instead focused on quantitating
known edit sites rather than predicting new sites. Variants were called from the
aligned reads with the GATK Haplotype caller and A-+G and T-C variants were
intersected with a set of 17,831 curated RNA-editing events from the DARNED and
RADAR RNA-editing databases. This left a set of 3,169 editing sites in 1,756 genes
where at least one editing event was called in a gene in a single sample. The functional effects of these events were annotated with Ensembl's Variant Effect Predictor
with the vast majority of edits occuring in the introns or regions flanking genes (Table
2.6).
Differential editing events between Synl-Tscl~/- and control mice were called
using a binomial test with a FDR cutoff of 0.1 and a fold change cutoff > 0.25,
2.3.
RESULTS
81
with the rationale that these events are more likely to be true events and to also be
biologically relevant. This process identified 32 differential editing events between
the Synl-Tscl-/- and control mice. These differential editing events are primarily in
introns and the 3' UTR of genes but do not overlap splice sites or known miRNA
target regions.
2.3.6
Syn1-Tsc1-/- mice have aberrant calcium signaling
Given that we observed differential expression of genes associated with calcium signaling pathways in the cortex of Synl-Tscl-/- mice (see Table 2.2 on page 76, we
hypothesized that knocking down Tscl may lead to an alteration of calcium signalling in neurons. We cultured cortical neurons from Synl-Tscl-/- and control mice
and transfected the cultures with the genetically encoded calcium sensor GCaMP3.
Figure 2-5 on the next page shows example images of increased calcium flux throughout the dendritic tree in Synl-Tscl-/- neurons. Synl-Tscl-/- neurons had more frequent bursts of calcium throughout the dendritic tree, spiking close to 300 % more
often (interspike interval:
35.0 seconds
13.5seconds t 3.4seconds in Synl-Tscl~/~ cells (n=5),
10.0 seconds in wild type cells (n=4), (p=0.047).
Previous work has shown deletion of Tsc2 causes overactive release of calcium
from the ER in non-excitable cells in tuberous sclerosis tumors[132], indicating that
increased calcium signalling may be a common feature of disrupting the mTor pathway.
Overactive serotonin signaling can cause autism-like symptoms in mice. [179]. In
addition, work on the same conditional Syn1-Tsc1-/- mouse model where the deletion
was restricted to only serotonin releasing neurons showed that Tscl deletion in those
neurons alone is sufficient to cause the autism phenotype[113] but not sufficient to
2.3.
RESULTS
82
CA)
10
pmn
-0
sec
3 sec
Figure 2-5: Cultured cortical Synl-Tscl-/- neurons have synchronized calcium flux throughout the dendritic tree (red arrows) whereas wild type neurons have more localized calcium
flux (purple arrows show decreased calcium). Neurons were transfected with GCaMP3 and
imaged at three second intervals. Synchronized calcium flux in the dendrites was much more
frequent in Syn1-Tsc1- neurons with an interspike interval of 13.5 seconds 3.4 seconds
(n=5) in Synl-Tscl-/- neurons and 35.0 seconds 10.0 seconds (n=-4) in wild type neurons
(p=0.047).
2.4. DISCUSSION
83
cause seizures. We hypothesized that overactive Htr2c receptor expression may be
the cause of the overactive calcium signaling in the Syn1-Tsc1-/- neurons.
2.3.7
Htr2c blocker halts aberrant calcium signaling
Cultured wildtype and Syn1-Tsc-/- neurons transfected with GCaMP3 were imaged.
Wildtype neurons had punctate calcium flux in the spines and not the shaft or soma
whereas the Syn1-Tsc-/- neurons had calcium bursts in the soma and long the entire
dendritic shaft.
Blocking Htr2c in the Synl-Tsc-/- cultures reversed the aberrant
calcium bursts in the shaft and soma and restored the punctate spiking at the spines
(Figure 2-6 on the following page).
2.4
Discussion
In this study we surveyed the transcriptional landscape of the cortex of mice lacking
Tscl in neurons.
To our knowledge, this is the first RNA-seq dataset of tuberous
sclerosis in any organism. We assayed not only differential expression at the gene
level but also splicing and editing differences caused by Tscl deletion. From this data
we identified a set of 253 differentially expressed genes, 167 genes with differential
splicing calls and a small set of 32 differential editing events when compared to mice
with the neuronal loss of Tscl.
Genes with differential transcription events were highly encriched for genes known
to interact or be affected by the mTor pathway. We also identified dozens of genes
differentially transcribed that are associated with autism, intellectual disability and
epilepsy, three of the strongest neurological phenotypes associated with tuberous
sclerosis. We identified several genes known to be involved in calcium flux in neu-
84
DISCUSSION
2.4.
84
2.4. DISCUSSION
U.......
777777=z~
0
~A
IWO~AAA
E
-- --
.
I
M illi
time
Figure 2-6: Data from a pilot experiment showing blocking the Htr2c receptor halts aberrant calcium spiking in Syn1-Tsc1-/- neurons. Treatment of cultured cells with SB24208,
an inhibitor of Htr2c, halts the rapid calcium spiking in cultured Syn1-Tscl-/ cortical cells.
GCaMP3 Each treatment is on the same cell and each trace represents the calcium signal in
a small random segment along the dendritic shaft. Treatment with Ro 25-6981, a selective
antagonist of NR2b does not halt the calcium spikes.
2.4. DISCUSSION
rons both via the MAP signaling pathway and through surface receptors known to
pass calcium across the membrane. We confirmed via calcium imaging that there
is increased an increased rate of calcium spiking in neurons lacking Tscl. The calcium spikes were not only more frequent and more diffuse, spreading throughout the
dendritic tree rather than at punctate spots at the synapses.
There is no cure for tuberous sclerosis and current therapies are focused on managing the myriad of symptoms caused by tuberous sclerosis. One promising avenue
is the treatment of low levels of rapamycin, an inhibitor of the mTor pathway to
alleviate the effects of overactivation of the mTor pathway. This treatment has been
shown to reduce autism symptoms and slow the growth of the cortical tubers in
both mouse models and patients. The long term effects of rapamycin are unknown,
however, and rapamycin treatment does not treat all of the symptoms or completely
ameliorate the symptoms it does treat. It may be necessary to treat the symptoms
of tuberous sclerosis with a cocktail of drugs. With this in mind, we used the set
of differential transcription events we identified as an atlas of potential sites of therapeutic intervention. We focused on Htr2c, as it was strongly upregulated in mice
lacking Tscl in neurons, has been shown to be be dysregulated in mouse autism models and passes calcium. We showed evidence that blocking Htr2c stops the diffuse
calcium activation in tuberous sclerosis neurons, suggesting that Htr2c is responsible,
at least in part, for the aberrant calcium activation in neurons lacking Tscl. Interestingly, Htr2c is a dense source of miRNA, seven different miRNA are encoded in its
introns[51]. Examining the affect of Htr2c upregulation on the expression of those
miRNA would be a useful follow up experiment.
One aspect of tuberous sclerosis that is not addressed by any current therapies is
the cortical disorganization that occurs in and around the cortical tubers. The cortical disorganization is due in part to improper migration during cortical development
85
2.4. DISCUSSION
86
and also due to a disorganized axonal arbor within and around the tubers. In the
mice lacking Tscl only in neurons, the tubers are not seen, yet the mice have severe
seizures. We identified a set of genes involved in axonal guidance and cell-matrix
interaction that may cause seizures indepent of cortical tubers. It is possible that
early intevention treating the expression of these genes could prevent some of the
seizure of intellectual disability phenotypes seen in humans.
Future work involves following up on the effect of upregulation of Htr2c in the
Synl-Tsc-1-/- knockout mouse to determine if an Htr2c antagonist could be a therapy
for the neurological symptoms of tuberous sclerosis. Several commercially available
Htr2c antagonists exist.
Risperidone, the first FDA approved compound for the
treatment the behavioral symptoms of autism, is an antagonist of Htr2c among
other receptors[111].
Further investigation into the effects of these antagonists on
the aberrant spiking in cultures and on the epilepsy and autism phenotypes of the
Synl-Tscl-/- mouse is warranted. In addition, it would be interesting to sequence
RNA from postmortem tuberous sclerosis patients to create a more comprehensive
catalog of differentially expressed, spliced and edited genes that may be involved in
the disorder.
Chapter 3
Transcriptome independent
retained plasticity of the
corticocollicular projection
of the mouse
3.1
3.1.1
Background
Superior colliculus
The superior colliculus in the mammal and the optic tectum in non-mammals integrates information from all visual areas in the brain and is important for the generation of saccades and orienting to stimuli[71]. The superficial layers of the superior
colliculus receive projections from the retinal ganglion cells and from the visual cor-
3.1. BACKGROUND
88
tex. The deeper layers of the superior colliculus receive projections from from many
other sensory areas of the brain including the ipsilateral auditory and association
cortex, the somatosensory cortex, the inferior colliculus and others. Thus, the superior colliculus is anatomically situated to be the locus of the sensory integration
involved identifying and orienting to salient stimuli from multiple senses. In addition
to receiving projects with information useful for orientation, the deep layers of the
superior colliculus project to motor nuclei in the brainstem controlling head and eye
orientation, further supporting the role of the colliculus as a site of sensory integration and subsequent motor output [73]. Information flows to the colliculus from the
various senses, computations are performed in the colliculus and the proper motor
behavior is output in the form of head movement and eye orientation.
Supporting this notion, ablating the superior colliculus of the primate causes a
permanent deficit in saccading to visual stimuli[4].
A study comparing lesions of
the frontal eye fields to lesions of the superior colliculus in the primate showed only
lesions of the superior colliculus affected saccades to visual stimuli[153], indicating
the superior colliculus is important for saccade generation. The superficial layers of
the superior colliculus are important for saccading to visual stimuli while the deeper
layers are important for saccading to non-visual stimuli. For example, ablating the
superior colliculus of the tree shrew, including the deep layers, results in a failure
to follow, track or even orient to any stimuli, including putative threats, whereas
ablation of only the superficial layers of the colliculus results in normal visually
guided behavior [33]. Similarly, deactivating the cat superficial superior colliculus by
cooling produces a lack of orientation to visual stimuli while leaving orientation to
auditory stimuli unimpaired[100]. The separation of the functions of the superficial
and deep layers of the superior colliculus are supported not only by lesion studies
but also the anatomy of the colliculus.
3.1. BACKGROUND
Anatomically, the superior colliculus is a highly organized structure with inputs
from the senses organized in topographic maps with respect to visual space and with
projections from different senses terminating in topographically similar positions. On
the surface of the colliculus, the topography of the contralateral retina is represented
as a high fidelity map of nasal retina in the anterior region and progresses through
the temporal retina in posterior regions[65]. The internal structure of the colliculus
is broadly divided into two major laminae: a superficial layer (sSC) comprising of
the stratum zonale (SZ), the stratum griseium superficiale (SGS) and the stratum
opticum (SO) and deep layer (dSC) comprising of the stratum griseum intermediale
(SGI), the stratum album intermediale (SAI), the stratum griseium profundum (SGP)
and the stratum album profundum (SAP) [160]. The superficial layers of the superior
colliculus recieve input from the retinae and the visual cortex [103] whereas deeper
layers receive input from other sensory areas such as the auditory cortex and the
somatosensory cortex [160]. This retinotopic organization becomes more diffuse in
the deeper layers of the colliculus compared to the high fidelity map of space in the
sSC. The connections within and between each layer of the superior colliculus are
also highly organized, with each layer containing a small, anatomically disparate set
of neurons.
In the sSC, there are five major cell types: marginal cells in the SZ,
horizontal and stellate cells in the SGS that process intralayer visual signals and
wide field vertical and vertical cells that integrate information from the superficial
and deeper layers of the colliculus. Visual information enters the sSC mostly from
the contralateral retina via the SO. This retinocentric information is combined with
information from the ipsilateral visual cortex via stellate cells and wide field vertical
cells. This visual information is sent to the deepest layers of the colliculus via the
vertical cells that combine it with information from afferents from other sensory
systems[80]. Neurons deep in the superior colliculus output to motor nuclei in the
89
3.1.
90
BACKGROUND
brainstem, eliciting eye movements and other orienting behavior. (Figure 3-1).
,O'
zWFV
Figure 3-1: Cell types of the superior colliculus. A schematic of the superficial visual
layers of the superior colliculus, with cell types labeled in their appropriate layers. Cells
marked in red are inhibitory and cells marked in black are excitatory. m = marginal, v =
vertical, s = stellate, h = horizontial, w = wide field vertical, SZ = stratum zonale, sSGS
- superficial stratum griseium superficiale, dSGS = deep SGS, zWFW = zone of widefield
vertical neurons. Used with permission from [136], Copyright Massachusetts Institute of
Technology.
The establishment and registration of the sensory specific maps in the layers of
the superior colliculus is important for its function and disruption of the topograph
map results in an impaired ability to locate and saccade to many sensory stimuli.
Quaia et al. showed that deactivating intermediate layers of the right superior colliculus with the GABA receptor blocker muscimol induced a reversible increase in
response latency, a decrease in speed and a decrease in accuracy to stimuli in the
left visual field. In addition this effect was strongest with stimuli in the zone of the
left visual field that was deactivated in the right colliculus[143].
More recently a
similar experiment was performed using optogenetic deactivation of neurons in the
3.1.
BACKGROUND
91
monkey superior colliculus. This allowed for a precise deactivation of subpopulations
of neurons in the superior colliculus and showed that saccade endpoints were shifted
away from the deactivated visual area[34].
3.1.2
Retinotopographic map formation in the colliculus
Topographic maps are an efficient way to represent information through maximization of wiring economy[38] and many sensory systems use a topographic map as a
first step of processing the information from the environment. The somatosensory
system employs a topographic map of the body in primary somatosensory cortex[1 16],
the auditory system has a frequency map in the cochlear nucleus [85] representing
the locations of resonant frequency on the basilar membrane and the visual system
has a topographic map of the retina projected on both the visual cortex and the
superficial layers of the superior colliculus.
The topographic map from the retina
to the superior colliculus is formed and refined to its mature state through several
gradual stages involving both activity dependent and independent processes.
The mature retinotopic map in the superior colliculus forms from projections of
the optic tract to the stratum opticum with nasal-temporal retinal ganglion cells
mapping to the anterior-posterior superior colliculus and dorsal-ventral retina mapping to lateral-medial superior colliculus. The question of how visual axons find the
correct zone to terminate in is an interesting one. A solution was suggested by Sperry
in the form of the chemotaxic hypothesis where multiple gradients of receptors on
the growth cones of axons afferent cells and gradients of ligands for those receptors
in the target neuropil could serve as signals for the axons to stop growing and to
elaborate terminal arbors[161]. This is an interesting hypothesis because it requires
only two orthogonal gradients of molecules to specify unique coordinates across the
3.1.
BACKGROUND
92
neuropil, much like a single integer and a single letter can uniquely specify every
square on a chessboard. The chemotaxic hypothesis was first borne out in the mammalian colliculus, with the initial, rough patterning of the map occuring prenatally
as the result of two main orthogonal gradients of molecular cues, the Eph receptors
and their ligands, the ephrins. EphA expressed on retinal ganglion cell (RGC) axons
interacts with ephrin-A expressed in the colliculus, acting to repel the axons. The
retinal ganglion cell axons enter the sSC from the anterior end, where ephrin-A is
less expressed and RGC axons with high EphA expression stop expanding and terminate. Continuing along the anterior-poterior axis of the colliculus, the gradient of
ephrin-A increases causing RGC axons with less and less EphA terminate[112].
A
similar process guides the lateral-medial mapping of the dorsal-ventral retina, using
gradients of the EphB receptor on RGC axons and ephrin-B expressed in the superior
colliculus. The utility of chemotaxic gradients is very robust and is involved in not
only the initial retinotopic map formation but in the registration of multimodal sensory maps on top of the retinotopic map in the colliculus. Similar gradients of these
same or closely related molecules have been identified in both warm and cold blooded
vertebrates in most sensory modalities. For example, a study from the Feldheim lab
looking at the somatosensory whisker map to colliculus sensory input showed using
knockout mice for ephrinA4 and ephrinA7 that gradients of ephrinA4 and ephrinA7
are necessary to register the somatosensory map on top of the retinotopic map in
the superior colliculus[173].
While the chemotaxic hypothesis explains how the intial rough retinotopic map
in the colliculus is formed it is does not completely explain how the mature, wellrefined topographic map arises. In addition to molecular cues the retinotopic map is
further refined through intrinsic spontaneous activity generated in the retina[188] [35]
and later via visual activation of the retina [29]. These two processes serve to fine-
3.1. BACKGROUND
93
tune the diffuse, exuberant elaborations generated through the initial chemotactic
mapping.
In rodents the immature retinae generate spontaneous bursts of infrequent activity which spread across the retinae and serve to synchronize the firing of neighboring
cells[188]. In mice this spontaneous activity begins at P6 and peaks at P10, petering
out at P14, after the eyes have opened[120]. These spontaneous waves have been
shown to be important for organizing multiple levels of the mouse visual system[l].
In the visual cortex it is well known that if this spontaneous synchronized activity is
disrupted through blocking of activity via retinal TTX injection[165] the formation
of ocular dominance columns is impaired. Similarly, in the colliculus, either blocking
spontaneous activity in the retina or inducing dissynchrony among inputs[112] results
in an abnormal retinotopic map formation in the colliculus. The work by McLaughlin is especially interesting- mice lacking the 32 receptor of the neuronal nicotinic
acetylcholine receptor lack retinal waves but have a normal amount of activity during
the first postnatal week, when majority of the 2nd order refinemnt of the retinotopic
map in the colliculus is occuring. #2-/- mice do not display a tight termination zone
by P8, with diffuse the retinocollicular connection being in approximately the correct
area but with a diffuse, unpruned arborization. This indicates that it is the pattern
of spontaneous activity in the retina, not the overall amount of activity that is important in the refinement of the retinocollicular connection, likely due to mechanisms
of strengthing synaptic connections by co-active inputs as suggested by Hebb[74].
Disruption of spontaneous activity affects not only the anatomy of the retinotopic
map map in the colliculus but also has a functional effect.
#2-/-
mice have abnor-
mal receptive fields in the colliculus but surprisingly cortical cells exhibit normal
receptive field size[184]. The same study showed that 02-/- mice perform poorly on
an optokinetic task tracking the mouse's head movement following horizontal and
3.1. BACKGROUND
94
vertical sinusoidal gratings. Such tracking has been shown to test the subcortical
visual centers[49]. Surprisingly, the same mice performed normally in a task measuring cortex-dependent spatial vision, suggesting that activity has a larger role in
determining the quality of the functioning ouput in the colliculus than in the visual
cortex.
The event of eye opening in rodents provides another form of activity that serves
to instruct the final refinement of the topographic maps and visual function in the
visual cortex and superior colliculus. Mice are born with their eyes closed until P13
%
but the eyelid is not completely opaque and is capable of transmitting up to 10
of available visible light [17].
This diffuse, attenuated light is capable of driving
visually responsive cells in the cortex and the superior colliculus one or two days
before eye opening. The light response through the closed eye is strong enough for
visual neurons to be able to discern secondary features of the light such as place,
orientation and even movement[93].
However, numerous studies in diverse verte-
brates have shown that this small amount of visual stimulation is not enough to
properly refine the cortical and subcortical visual connections and highly patterened
activation after eye opening is very important for the functional development of
the visual system in both humans and rodents. The visual neuropil requires light
evoked activity within a critical window called the critical period in order to develop
properly; if deprived of light evoked activity during this critical period the visual
system development is retarded in a profound fashion. The critical period for the
visual system ranges in species, for months in rodents to several years in humans
[16]. Studies of human infants with uncorrected cataracts, where a cloudiness of the
lens prevents most extrinsic light from accessing the retina, is an excellent model
for studying the functional effects of early visual visual activity in humans. Lack
of extrinsically driven visual activity has a profound effect on visual system func-
3.1.
BACKGROUND
tion and children born with cataracts left uncorrected for 7-9 years are functionally
blind even after cataract correction[ 110]. Even brief periods of early deprivation of
patterened activity cause defects in lower-order processes such as spatial acuity, orientation selectivity, directional selectivity[22] and higher order cognitive processes
such as discernment of implied forms and depth based on visual cues[59]. Subsequent
visual exposure after the critical period has passed is unable to correct these deficits.
In cats, lack of visual activity through dark rearing or lid sutures for several months
induces gross impairments in visuomotor and visual acuitiy[122]; in monkeys dark
rearing causes a suppression in the naso-temporal direction as well as a deprivation
induced nystagmus[176].
In addition to functional effects on behavior, light deprivation causes many anatomical and electrophysiological changes in the visual system neuropil both in the visual
cortex and the superior colliculus. Binocular activation of the eyes is necessary for
proper development of binocular cells in the visual cortex; depriving an eye of visual experience removes the ability of the deprived eye to drive the binocular cells[9].
Furthermore, long term deprivation through dark rearing decreases the responses
of single units in the visual cortex to orientation and direction[57], and causes the
receptive field size of neurons in the visual cortex to remain immature[56]. Visual
experience in the mouse is necessary for proper retinocollicular projection formation
and reduction in visual experience through dark rearing reduces the density of fibres
from the retina to the stratum opticum in mice[142]. However, unlike the visual cortex, early visual experience is not necessary for initial receptive field refinement in
the superior colliculus but rather visual experience in adulthood past P60 is required
for the maintenance of the receptive field size[28]. This finding was supported by
work showing that dark rearing affects spatial tuning of single neurons in the superior
colliculus, but not their orientation or direction selectivity or receptive field size[182],
95
3.1. BACKGROUND
even with long term deprivation up to P60. Early exposure to light prevents the
destabilization of receptive fields in the colliculus due to long term deprivation in
the adult, even though the receptive fields of both the dark reared and light reared
animals refine to the same size[31]. This indicates that even relatively brief periods
of light exposure can have long-term effects on the colliculus that are not necessarily reflected in the first-order functional or anatomical characteristics. These more
subtle second-order effects could be molecular or transcriptional modifications in the
neuropil of the superior colliculus.
Evidence for molecular modifications due to eye opening are strong in both the
visual cortex and the superior colliculus, especially in the period following eye opening, though most work has been done in the visual cortex. Early studies found that
blockade of the NMDA receptor (NMDAR) during the period of eye opening results
in disruption of normal ocular dominance column formation and orientation selectivity in the visual cortex[14]. NMDAR activation fluxes Ca 2 - into the cell, activating
cAMP which in turn activates CREB leading to activation of the transcription factor
CRE and transcription, raising the possiblity that visual activity resulting in activation the of NMDAR changes the transcriptional or molecular state in the visual
system neuropil. Visual manipulations have confirmed this in the cortex, showing
that brief visual experience induces the gene expression of a host of the intermediate
early gene transcription factors[123]. In addition to activating signal transduction
pathways downstream of the NMDAR, light exposure changes the composition of
the NMDAR itself, inducing a subunit switch from predominantly NR2B to predominantly NR2A in both the visual cortex[27] and the superior coliculus[18]. In the
superior colliculus, early light exposure also decreases phosphorylation of the NR2A
receptor, which reduces the decay time of the NMDA current mediated by NR2A[168]
and also affects the distribution of the NR2A receptor at the synapse, moving it to
96
3.1.
BACKGROUND
the center of the synapse[169].
The observation that the NMDA receptor is a crucial mediator of light-induced
changes in the visual areas and that exposure to light causes activation of several
intermediate early gene transcription factors indicate that light exposure may cause
the neuropil to enter a new state through modification of the the transcriptional
program in the visual neuropil. The advent of relatively inexpensive transcriptome
profiling procedures has made performing analyses regarding which transcripts may
be differentially regulated feasible. One study showed that dark rearing from birth
to P27 or monocularly depriving the animal causes the differential regulation of an
thousands of transcripts in the primary visual cortex of mice[174]. Short term monocular deprivation of only four days differentially regulated more transcripts than long
term deprivation, indicating that there may be compensatory mechanisms that occur with long term deprivation[174]. The differentially regulated transcripts included
GABA receptors, NMDA receptors and AMPAR receptors but not metabotropic glutamate receptors indicating the excitatory and inhibitory circuits in V1 are regulated
with light deprivation. Interestingly, the insulin growth factor (IGF) pathway was
shown to be upregulated after monocular deprivation and IGF1 injection prevented
the ocular dominance shift caused by monocular deprivation, giving a role for the
IGF pathway in ocular dominance column formation. A different study using monoenculeation instead of monocular deprivation showed that 4 days of ME regulates
a small set of tens of genes in the visual cortex of mice, mostly IEG transcription
factors noted in previous, PCR-based studies [106][123]. It is hard to reconcile the
results of these two studies, as there is very little overlap between the two gene sets
identified as differentually regulated between the two studies, and one study finding
two orders of magniude more differentually regulated genes than the other, despite
very similar methods[175]. It is possible that CRE activation through ERK has not
97
3.1.
BACKGROUND
98
had time to ramp up transcription regulation downstream of the IEGs[24]. Nevertheless, transcriptome profiling of the visual cortex has been fruitful, demonstrating the
importance of the IGF pathway in ocular dominance column formation[174]. Multiple studies have confirmed, looking at single genes, differential regulation of several
individual transcripts in the visual cortex through deprivation[90, 131] including a
microRNA[115, 167].
In addition to regulation at the gene level a study showed
that visual experience induces differential isoform expression of TrkB in the visual
cortex[20]. These studies show that light activation of the retina can induce both
transient and long term changes in the molecular composition of the synapse in the
superior colliculus and the visual cortex.
Long term light deprivation also induces profound anatomical changes in the
projection from the visual cortex to the superior colliculus. Eye closure until P16,
three days after eye opening results in corticocollicular axons being stripped of their
terminal arbors in the superfificial layers of the superior colliculus. In addition to
changes in the axons themselves from the visual cortex, the axons of the visual cortex
cause filopodia induction and new synapse formation in the superficial layers of the
superior colliculus.[137] Thus eye opening and the onset of patterened activiation
induces changes in the function, molecular composition and anatomy of the superior
colliculus. Similar work showed these effects were reversible by opening the eyes[64]
(Figure 3-2 on the following page).
Despite receiving direct input from the retina and the visual cortex, and evoked
activity causing well documented alternations in the molecular composition and
anatomy of the colliculus there have been no transcriptome-wide studies of the effect
of light deprivation on the superior colliculus. Furthermore, almost no work has been
done examining transcriptome-wide splice variant changes in any visual areas of the
brain due to eye opening. Splicing differences are important to look at in the brain as
3.1.
3.1.
99
BACKGROUND
99
BACKGROUND
BEO
A
P13
EC
EO
B
F
C
G
G
P15
P16
D
-~_______
P17
I
E
P19
A
P
A
P
Figure 3-2: Eye opening refines the corticocollicular projection. Tracings of the corticollicular projection show laid out along the anterior-posterior axis show eye opening induces
elaboration in the terminal zone and pruning of exuberant branches along the trunk of the
axon. Eye closure induces a persistent pruning of the connections along the entire length of
the axon. Used with permission from[64], copyright Massachusetts Institute of Technology.
3.2. METHODS
100
alternative splicing levels area highest in the brain compared to the rest of the body
and have been shown to have functional effects in some areas of the brain[21][1211.
We chose to further investigate the effects of eye opening on the superior colliculus
by examining transcriptional changes in the superficial superior colliculus for much
longer periods of deprivation but before the end of the critical period. We hypothesized that long term deprivation may result in the neuropil of the superficial superior
colliculus being transitioned to a new transcriptional state, and this new state is in
part what causes light induced retina activation to have such profound effects on the
molecular composition and anatomical structure of the superior colliculus.
3.2
3.2.1
Methods
Eye closure manipulation and colliculus dissection
Sprague Dawley rats were ordered from Charles River Laboratories and were kept in
a clean facility on a 12 hour day/night cycle. The male Sprague Dawley rats in each
litter were divided into eyes closed and eyes opened groups. At P12 to P13, the days
prior to natural eye opening in the rat, the rats were brought down to our laboratory
to perform eye-suturing of the eyes closed group. Rats of the eyes closed group were
anesthetized with inhaled isoflurane first applied in a bell-jar and once breathing
had slowed, delivered in a controlled fashion through a vaporizer attached to a a
nose-cone. We sututed the eyelids closed with black 6-0 monofilament nylon suture
using an interrupted stitch with at least five stitches per eyelid. On top of the suture
we applied Mastisol@ (Ferndale #0523),
a translucent liquid adhesive adhesive to
make it more difficult for the rats to damage and remove the suture. On top of the
suture and the glue we applied Emla@ cream, a topical lidocaine-based analgesic to
3.2. METHODS
101
mitigate pain. We checked the sutured animals health status every day for four days
after the surgery and observed no sign of distress in any of the animals. Every two
days after P14 we reapplied the glue and after P18 we applied the glue every day
as deemed necessary by visual inspection of the sutures. Any rat with the sutures
removed was discarded from the study. Rats from the eyes open group and the eyes
closed groups from the same litter were anesthetized the day of suturing but were
not sutured or glued and Emla@ cream was not applied. These procedures followed
MIT IACUC-approved protocols.
At P20 we sacrificed both the eyes closed and the eyes open groups of animals and
performed a dissection of the superior colliculus in a manner enriched for the superficial layers of the colliculus (Figure 3-3 on the next page). Rats were anesthetized
with isoflurane in a bell jar and immediately decapitated and the head placed on dry
ice.
The brain was dissected on ice with tools treated with RNaseZap@ (Ambion
#AM9780)
to protect against RNA degradation by RNases. The skull was removed
and the brain was washed with ice-cold RNase free PBS prior to dissection of the
colliculus.
The cortex was removed and the superior colliculus was dissected and
immediately transferred to 1 mL of RNALater@, a solution designed to inactivate
RNAses and stabilize RNA to prevent degradation. Dissections were reproducible
and uniform, representing a superficial-enriched dissection of the superior colliculus.
RNALater treated colliculi were kept at 4C for 24 hours and were then transferred
to -20'C for indefinite storage.
3.2.2
Total RNA isolation and quality control
The superior colliculus was removed from the RNALater@ and the and placed in 1
mL Qiazol@ (Qiagen #79306)
reagent. The colliculus was immediately homogenized
3.2.
102
102
METHODS
3.2. METHODS
SZ
Multimodal
Motor SC
/
Visual SC
RNAseq
dissection
psth
Figure 3-3: A schematic of the superior colliculus dissection, viewed as an idealized, sagittal slice. The superior colliculus was dissected to enrich for the superficial layers of the
superior colliculus, capturing the stratum zonale (SZ), stratum griseum superficiale (SGS)
and the stratum opticum (SO). Also captured with this dissection was a small amount of
the intermediate layers, the stratum griseum intermediale (SGI) and the stratum album
intermediate (SAI) which recieve multisensory inputs.
3.2.
METHODS
103
using a Tissue TearorTM (Research Products International Corp #985371)
minutes or until the sample appeared to be completely homogenized.
for two
0.2 mL of
chloroform was added to the sample and the sample was vortexed for 15 seconds
until the solution was uniform. The sample was centrifuged for 25 minutes at 4C
and the aqueous layer was saved and added to 0.5 mL ice-cold isopropanol with 1 P1
of glycogen to help visualize the RNA pellet. The sample was centrifuged for 15
minutes at 4 C and the supernatant was discarded. The pellet was washed with
75 % ethanol and recentrifuged for 15 minutes at 4 'C. The supernatant was poured
off and the pellet was allowed to air-dry inverted for 10 minutes at room temperature.
After air-drying the pellet was redissolved in 48 Ill DEPC-treated H20 of Ix DNase
buffer (Roche #04-716-728-001).
To clean up the isolated total RNA, we treated the RNA with 1 pIl DNase 1
(ten units) (Roche) and 1 pl RNase inhibitor (ten units) for 20 minutes at room
temperature to remove the genomic DNA. We then used the RNeasy@ MinElute@
cleanup kit (Qiagen #74204) to remove the enzymes, leftover DNA, salt and other
contaminants from the isolated total RNA following the manufacturers instructions
except for washing the RNA with 80 %ethanol three times instead of once. total RNA
was eluted in 30 pl of RNAse free H 2 0 and stored at -80 'C until library creation.
This cleanup step removes all RNA fragments that are less than 200 nucleotides
in length, an important point to note for downstream analyses as there are many
annotated small non-coding RNAs which fall into the group of RNAs under 200
nucleotides.
The purity of total RNA was assayed using a NanoDrop spectrophotometer
(Thermo Scientific). The absorbance at 230, 260 and 280 ratios were measured from
1.5 pl of sample and samples with 260/280 ratios less than 1.8 and 260/230 ratios
less than 2.0, indicating the presence of protein, phenol or other contaminants, were
3.2.
METHODS
104
re-phenol extracted and purified as above. Samples not conforming to the 260/280
and 260/230 ratios after reextraction were discarded and their sister sample from the
litter pair was also discarded. 2pl of the samples were then run on a Agilent 2100
Bioanalyzer to estimate the concentration and degradation of the isolated total RNA.
Samples with RNA integrity numbers less than 9, indicating degradation of the RNA
were discarded along with their sister sample. Samples were stored at -80 C until
library creation.
3.2.3
cDNA Library creation and sequencing
cDNA libraries for paired-end sequencing on the Illumina Genome Analyzer II were
prepared following the Illumina protocol (#1004898 Rev. D) with a few alterations.
Briefly, mRNA was bead-purified from total RNA using Sera-Mag poly-T oligo attached beads and 1 pl of purified mRNA was saved and run on an Agilent Bioanalyzer
to check for yield and RNA integrity. RNA was fragmented for 2 minutes at 94 C,
converted to cDNA, end repaired and the 3' ends were adenylated. Adaptors were
ligated and the cDNA was run on a 2 % agarose gel. The band at
was cut out and purified.
200 base pairs
The resulting product was PCR amplified for 10 cycles
using primers against the Illumina adaptors.
The final library was run on an Agi-
lent Bioanalyzer to confirm proper size selection. A total of six samples, three eyes
open control samples and three eyes closed samples were submitted to the BioMicro
Center at MIT for sequencing on the Illumina Genome Analyzer II with 36 base pair
paired-end reads.
3.2.
METHODS
3.2.4
105
Informatics analysis
We processed the RNA-sequencing data using the RNA-seq pipeline implemented in
version 0.7.9a of the bcbio-nextgen analysis package. Briefly, we trimmed off poor
quality ends with AlienTrimmer[43], using a cutoff of phred[551 score of 5 or less[105]
and trimmed portions of reads and anything after it matching the first 13 bases of the
Illumina universal adapter sequence to remove read-through contamination caused
by the read length being longer than the insert size for a fragment. We also trimmed
polyA and polyT homopolymer sequences from reads the 5' ends of reads.
A STAR[48] index was created from a combination of the rattus norvegicus
genome, build rn5 and the Ensembl release 74 gene annotation. Reads were aligned
with STAR version 2.3.14z using the default settings, with the exception of filtering
out reads mapping more than 10 times to the genome. Counts of reads mapping to
genes in the Ensembl annotation were calculated using FeatureCounts[97].
Quality
metrics including mapping perecentage, rRNA contamination, coverage, read quality,
adapter contamination and others were calculated using a combination of FastQC,
RNA-SeQC[45] and custom functions using the bcbio-nextgen and bcbio.rnaseq
packages. Chapter 3 of this work has more details.
3.2.5
Differential expression
Differential gene expression was performed with DESeq2 version 1.4.1 on using R
version 3.1.0 using a blocked design taking into account litter and deprivation status of the mouse.
DESeq2 was chosen based on simulations which showed it has
the best performance out of several popular algorithms tested for simple two-factor
comparisons and two-factor paired comparisons. Using data simulated to have the
same distribution as our count data, we show that these procedures can expect to be
3.2.
METHODS
106
fold change
sensitivity
specificity
2+
0.96
0.95
1.5-2.0
0.37
0.75
1.1-1.5
0
0
1.05-1.1
0
0
Table 3.1: Power estimation of the eyes open vs. eyes closed experiment at a range
of fold changes. Simulating a RNA-seq experiment using a similar transcript abundance
distribution, sample size and biological variation shows that an experiment similar in size
to the colliculus experiment should be expected to pick out fold changes twofold or greater
reliably but fail to detect more moderate fold changes. For each range of fold changes
the true positive rate (sensitivity) and false positive rate (specificity) are calculated. Fold
changes greater than two-fold have high sensitivity and specificity but anything below that
is likely to be missed for an experiment of this size.
able to pick out a large percentage of the highest simulated fold changes. However
we will miss most moderate changes with our experimental setup (Table 3.1). See
Chapter 3 for a description of the simulation with justification for choosing DESeq2
over the other differential expression callers.
Exon level counts were produced and analyzed using DEXSeq version 1.10.8 using a similar model design as the gene-level differential expression.
Differentially
expressed exons were called with an FDR cutoff of 0.1 and filtered for exons with an
absolute fold change of 1.5 or greater, since simulations show that this is the limit
of detectability for an experiment of this size.
3.2.6
Corticollicular projection mapping and quantitation
Projections from the visual cortex to the superior colliculus in rats were labelled and
reconstructed as described in [64]. Briefly, cortical afferents to the sSC of rats were
3.3.
RESULTS
107
labelled with DiI soaked gelfoam inserted below the pial membrane over the visual
cortex from P17-P20. After two days of gel foam placement, rats were anesthetized
and perfused with 4% PFA and post-fixed for 24 hours.
150-200 Pm thick slices
were cut from the sagittal plane of the neocortex and midbrain and sections were
collected and mounted with Fluoromount. Stacks of images were collected on a Nikon
PCM2000 (MVI) confocal scanning microscope using a 40X objective, taking 1.5 Jrm
z-steps. Stacks of images were flattened and compressed into a single 2D projection
of the 3D stack. The colliculus was divided into five quadrants from the anterior
to posterior end and the number of branch points in each quadrant was counted
as a measurement of arborization.
Example images from this process appear in
previous work[64], here we quantitate the entire dataset from that work, including
the EC-÷EO reopening an EO-÷EC closure. EC-+EO rats had their eyes closed
from P13-P19 and were opened for two days from P19-P21. EO-4EC rats had their
eyes opened from P13-P19 and were closed for two days from P19-P21.
3.3
3.3.1
Results
Corticocollicular projection remodelling
As reported previously[64], eyelid suturing until P20 has a profound effect on the
axonal branching of the corticocollicular projection in the colliculus. Normal eye
opening results in exuberant branching of the primary axon from the visual cortex
to the superficial layers of the superior colliclus; suturing the eyes closed until P20
results in a pruning of the exuberant branches. We quantitated the effect of inducing
pruning by closing the eyes of a normal animal or reversing pruning by opening the
eyes of an eyes-closed animal on the degree of branching of axons in the corticotectal
3.3.
RESULTS
108
projection to the superficial superior colliculus (Figure 3-4 on the next page). An
analysis of variance (ANOVA) shows significant variation (p < 0.01) among the eye
state groups. A post hoc Tukey test showed manipulating the eye state from the
normal eye opening (EO) causes significant differences when the eyes are closed (EC)
(p < 0.01) but not when the eyes are initially closed and then opened (ECtoEO). This
supports the conclusion of [64] that reopening the closed eyes restores the structure
of the corticocollicular projection to its normal state.
3.3.2
Sequencing quality control
Sequencing libraries contained 50-70 million reads per lane and about 80 % of reads
mapping to known genes using the Ensembl release 74 of the rat annotation. After
trimmed mean of M-values (TMM) normalization[149], samples had a similar readcount distribution across genes (Figure 3-5 on page 110. Samples had a high Pearson
correlation to each other and clustered together based more on the litter they were
drawn from, not the experimental condition (Figure 3-6 on page 111).
Observing
clustering by litter validates the choice to litter-match the samples; without litter
matching the major component of the variation between the samples would be unable
to be corrected for.
3.3.3
Differential gene expression
The MA-plot (Figure 3-7 on page 112) of closed vs. open samples shows there is are
only moderate fold changes in genes between the eyes closed and eyes open state,
where the eyes open animals have had five days of eye opening.
A small set of genes that were flagged as differentially expressed is shown in
Table 3.2 on page 113. Interestingly two of these genes, Mt-nd4l and Mt-nd6, code
109
109
3.3. RESULTS
3.3. RESULTS
SW?
s
tss-'t ~
40-
20+
04
4rz~z~~
_
-o
_____
401J
[
________
1
1---
________
.1
____
-a
20-
--
-
*I
1
2
3
4
5
1
I
I
I
I
2
3
4
5
quadrant (anterior to posterior)
Figure 3-4: Effect of eye closure state on corticocollicular arbor development. Raising rats
with eyes sutured closed from P13 until P19 (EC) prevents the exuberant axonal branching
of the corticollicular projection from the visual cortex to the posterior superficial superior
colliculus compared to the mouse with normal eye opening (EO). Reopening the eyes for
two days at P19 (ECtoEO) restores the EO corticocollicular axonal topography. ANOVA
with a post hoc Tukey tests shows manipulating the eye state from the normal eye opening
(EO) causes significant differences when the eyes are closed (EC, p value < 0.01) but not
when eyes are closed and then reopened (ECtoEO).
110
3.3. RESULTS
10000-
100-
2e.07 7
A~c
Aanope
B-ckwd
(a)
Bjope-
C
c
C_ po
Acksed
Ape
B
d
Bopen
CAc
C-Ope
(b)
Figure 3-5: a) The libraries have a varying number of total mapped read counts. b)
Trimmed mean of M-values (TMM) normalization[149] effectively normalizes gene expression. Samples have a (litter) (closed/open) naming scheme.
3.3.
RESULTS
ill
111
3.3. RESULTS
A_dosed
4000
2000
-
----
-
+------------+-
-
-
Aopen
CJ
0
0
C
oPen
0open
2000
B_closed
C_dosed
4000
10000
0
10000
PCi 80%
Figure 3-6: Clustering of gene expression in the rat colliculus. Multidimensional scaling
(MDS) estimates the similarity between high-dimensional samples. MDS loosely clusters
the samples into pairs based on litter (A, B and C), justifying including litter in the
differential expression model. Samples have a (litter)_(closed/open) naming scheme.
3.3.
RESULTS
-
rC4
112
0
4**.....
-
* ~
V
0
'.5
0
0
(.4-
0.1
10.0
1000.0
100000.0
mean expression
Figure 3-7: MA-plot (mean expression vs log2 fold change) between eyes open and eyes
closed rats. The fold change between the two conditions is very modest. Points colored
red are significantly different with a FDR < 0.1. Positive log2 fold changes are genes that
are more highly expressed in the eyes closed compared to the eyes opened animals.
3.3. RESULTS
113
gene id
symbol
log 2 ( cloed)
FDR
ENSRNOG00000011292
Colla2
0.4540360
4.087336e-03
ENSRNOG00000033299
Mt-atp8
-0.6846459
1.200520e-08
ENSRNOG00000031053
Mt-nd4l
-0.3978127
3.562899e-03
ENSRNOG00000029042
Mt-nd6
-0.6556634
6.030092e- 13
Table 3.2: Differentially expressed genes with with FDR < 0.05. A model was fit using
DESeq2, taking into account the within-litter variation and calling difference between eyes
closed and eyes open animals. Positive log2 fold changes are genes that are more highly
expressed in the eyes closed compared to the eyes opened rats. Mt-atp8, Mt-nd4l and
Mt-nd6 are all involved in the electron transport chain.
for NADH dehydrogenase, an enzyme which starts the electron transport chain by
catalyzing the oxidization of NADH by ubiquinone[5]. A third protein, Mt-atp8 codes
for mitochondrial ATP-synthase, another protein involved in the electron transport
chain.
3.3.4
Possible X-linked cofactors in LHON
Previous work has shown eye opening to increase the energy load on the visual system
neuropil[189]. Loss of function mutations in either Mt-nd4l or Mt-nd6 causes Leber's
hereditary optic neuropathy (LHON)[26]. LHON causes the selective death of RGCs
in the fovea.
An interesting feature of LHON is that although it is primarily a mitochondrial
disease there is a much higher prevalence in males and an incomplete penetrance in
affected families. One possible explanation for this phenomenon is that there may be
a cofactor on the X chromosome that is also mutated in affected male individuals[76],
[82],[155]. Using this additional information we looked at genes on the X chromosome
3.3.
RESULTS
114
which did not reach significance but had a confidence interval highly skewed in one
direction. We identified several genes meeting these criteria, including genes involved
in the electron transport chain (Cox7b) actin dynamics (Tmsbx, Tsmbll, Smarcal)
and oncogenes (Rab9b, Mctsl).
3.3.5
Differential exon expression
We found 212 differentially expressed exons at FDR < 0.1, but none differentially
expressed at levels 1.5 fold or greater.
(Figure 3-8 on the following page) shows
very clearly that these samples are highly uniform even when considering individual
exons.
3.3.6
Dopamine receptor expression in the superior colliculus
Evidence for the expression of the dopamine receptor Drdl in the superior colliculus
is mixed. Previous work looking at Drdl, Drd2, Drd3 and Drd5 dopamine receptor mRNA expression in the rat CNS found evidence only of Drd2 in the superior
colliculus[108].
Another study looking at Drdl and Drd2 expression found only
Drd2 in the superior colliculus[187]. An earlier study looking specifically for Drdl
found no Drdl mRNA but did find the presense of the Drdl protein, as it bound
a radiolabeled Drdl antagonist[107].
In this work we show that the composition
of dopamine receptor mRNA is much more diverse than previously reported. We
detect Drdl, Drd2, Drd3 and Drd5 at low levels in the superficial superior colliculus
(Figure 3-9 on page 117). tdTomato expressing Drdl neurons and EGFP expression
in Drd2 neurons show a segregation of Drdl and Drd2 into distinct layers of the
superior colliculus. Drdl is expressed in the all three sublaminae of the superficial
3.3. RESULTS
115
UI)
0)
Ci
o0
-o
0
0)
0.1
10
1000
10000
mean expression
Figure 3-8: MA-plot of differential exon analysis. Fold changes in exon usage are even more
moderate than the gene-levels changes. Although exons are called differentially expressed,
they are at a very low fold change (red points). These are likely to be false positives so
we discarded them. Positive log2 fold changes indicate exons expressed more highly in the
eyes closed samples. Noisy exons with low expression were filtered out from the analysis.
3.4.
DISCUSSION
116
superior colliculus and Drd2 is expressed in intermediate layers of the superior colliculus. There is sporadic overlap between Drdl and Drd2 positive cells, mostly in
the intermediate layers of the superior colliculus (Figure 3-9a on the following page).
3.4
Discussion
This work shows that despite changes in protein translation immediately after eye
opening[169] and persistent structural changes to the corticollicular projection with
long term deprivation[64]
[137] we do not detect many persistent changes in gene ex-
pression in the superior colliculus with prolonged visual deprivation. Previous work
done with similar sample sizes in the visual cortex[174] and [106] found hundreds of
differentially expressed genes in the visual cortex under a similar prolonged light deprivation paradigm. These results and post-hoc power calculations of this experiment
support the conclusion that maintenance of the structural changes of projections to
the superior colliculus with eye opening do not depend on changes in gene expression
to be maintained.
Deprivation on the order of several months reduces the number of projections
from the stratum opticum of the rat[141] and reduces the size selectivity [145] and
receptive fields of superior colliculus neurons[30]. Superior colliculus receptive field
formation is activity-independent[28] whereas the receptive field formation in the
visual cortex is activity-dependent[138]. The period we assayed in this experiment
falls in between the early chemotaxis-dependent formation of the retinotopic map in
the superior colliculus and the late activity dependent maintenace of the retinotopic
projection. This work combined with other work is evidence for a process where the
initial map is formed via chemotactic cues, after which eye opening causes brief, short
term changes in gene expression leading to structural changes in the corticocollicular
3.4.
3.4.
117
117
DISCUSSION
DISCUSSION
D
(
D2
(a)
4-
2-
0Drdl
Drd2
Drd3
Drd4
Drd5
(b)
Figure 3-9: Dopamine receptor subtypes are segregated into distinct layers of the superior
colliculus. a) Drdl tdTomato neurons are enriched in the superficial superior colliculus and
Drd2 EGFP positive neurons are expressed in the SO and SGI of the superior colliculus[19].
b) RNA-seq shows expression of Drdl, Drd2 and Drd5 with possible light expression of
Drd3 (Figure 3-9b). FPKM is Fragments Per Kilobase of transcript per Million mapped
reads.
3.4.
DISCUSSION
projection to the superior colliculus. After eye opening the activity genes involved
in structural plasticity return to steady-state, barring a set of mitochondrial genes
involved in handling the increased energy load on the superior colliculus caused by
the influx of signal to the colliculus from the retina. Long term deprivation on the
order of months induces a second round of activity-dependent structural changes in
the colliculus leading to functional deficits.
We confirmed the finding that light driven retinal activity causes an increase in
energy load on the colliculus and identified a small set of genes which may be coregulated with the mitochondrial genes on the X chromosome. Mctsl has been shown
to have anti anti-apoptotic effects; it is possible that there are additional mutations
in Mctsl or Cox7b that exacerbate the mitochondrial mutations in affected individuals. Most LHON studies have focused on sequencing the mitochondrial genome and
it would be interesting to perform exome sequencing on families with affected and
unaffected individuals and see if any variants on the X chromosome segregate with
affected status in this set of candidate genes.
Finally we showed strong expression of the dopamine receptors Drdl and Drd2
in the superior colliculus, resolving a conflict in the literature about whether or not
Drdl is expressed in the colliculus. We observed not only expression of Drdl but a
segregation of Drdl and Drd2 into two distinct zones laminae, with Drdl expressed
mostly in the superficial superior colliculus and Drd2 in the intermediate superior
colliculus.
This observation lead to a work assessing the functional significance of
this uncharacterized dopamine projection to the superior colliculus [19].
118
Bibliography
[1] James B Ackman, Timothy J Burbridge, and Michael C Crair. Retinal waves
coordinate patterned activity throughout the developing visual system. Nature,
490(7419):219-225, October 2012. 93
[2] Enis Afgan, Brad Chapman, Margita Jadan, Vedran Franke, and James Taylor.
Using cloud computing infrastructure with CloudBioLinux, CloudMan, and
Galaxy. Curr Protoc Bioinformatics, Chapter 11:Unitll.9-11.9.20, June 2012.
22
[3] Gael P Alamancos, Eneritz Agirre, and Eduardo Eyras. Methods to study splicing from high-throughput RNA sequencing data. Methods Mol Biol, 1126(Chapter 26):357-397, 2014. 40, 42
[4] J E Albano and R H Wurtz. Deficits in eye position following ablation of
monkey superior colliculus, pretectum, and posterior-medial thalamus. J Neurophysiol, 48(2):318-337, August 1982. 88
[5] Bruce Alberts, Dennis Bray, Karen Hopkin, Alexander Johnson, Julian Lewis,
Martin Raff, Keith Roberts, and Peter Walter. Essential Cell Biology, Fourth
Edition. Garland Science, October 2013. 113
[6] Simon Anders and Wolfgang Huber. Differential expression analysis for sequence count data. Genome Biol, 11(10):R106, 2010. 32
[7] Simon Anders, Alejandro Reyes, and Wolfgang Huber. Detecting differential
usage of exons from RNA-seq data. Genome Res, 22(10):2008-2017, October
2012. 40, 66
[8] Charles Anderson. Docker [Software engineering]. IEEE Software, 32(3):102c3, 2015. 22
BIBLIOGRAPHY
[9] A Antonini and M Stryker. Rapid remodeling of axonal arbors in the visual
cortex. Science, 260(5115):1819-1821, June 1993. 95
[10] Benjamin D Auerbach, Emily K Osterweil, and Mark F Bear. Mutations
causing syndromic autism define an axis of synaptic pathophysiology. Nature,
480(7375):63-68, December 2011. 59, 62
[11] Brenda Bass, Heather Hundley, Jin Billy Li, Zhiyu Peng, Joe Pickrell, Xinshu Grace Xiao, and Li Yang. The difficult calls in RNA editing. Nat Biotechnol, 30(12):1207-1209, December 2012. 80
[12] Helen S Bateup, Caroline A Johnson, Cassandra L Denefrio, Jessica L Saulnier,
Karl Kornacker, and Bernardo L Sabatini. Excitatory/Inhibitory Synaptic Imbalance Leads to Hippocampal Hyperexcitability in Mouse Models of Tuberous
Sclerosis. Neuron, 78(3):510-522, May 2013. 55
[13] Anna Bauer-Mehren, Michael Rautschka, Ferran Sanz, and Laura I Furlong.
DisGeNET: a Cytoscape plugin to visualize, integrate, search and analyze genedisease networks. Bioinformatics, 26(22):2924-2926, November 2010. 76
[14] M F Bear, A Kleinschmidt, Q A Gu, and W Singer. Disruption of experiencedependent synaptic modifications in striate cortex by infusion of an NMDA
receptor antagonist. J Neurosci, 10(3):909-925, March 1990. 96
[15] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a
practical and powerful approach to multiple testing. J Roy Stat Soc B Met,
57(1):289-300, 1995. 66
[16] N Berardi and T Pizzorusso. ScienceDirect.com - Current Opinion in Neurobiology - Critical periods during sensory development. Curr Opin Neurobiol,
2000. 94
[17] Andrew Bierman, Mariana G Figueiro, and Mark S Rea. Measuring and predicting eyelid spectral transmittance. J Biomed Opt, 16(6):067011, June 2011.
94
[18] K E Binns and T E Salt. Developmental changes in NMDA receptor-mediated
visual activity in the rat superior colliculus, and the effect of dark rearing. Exp
Brain Res, 120(3):335-344, May 1998. 96
120
BIBLIOGRAPHY
[19] Andrew Bolton, Yasunobu Murata, Rory Kirchner, and Martha ConstantinePaton. A diencephalic dopamine source projects to the superior colliculus,
where D1 and D2 receptors segregate to behaviorally relevant zones. Cell Rep.
117, 118
[20] Bethany K Bracken and Gina G Turrigiano. Experience-dependent regulation
of TrkB isoforms in rodent visual cortex. Dev Neurobiol, 69(5):267-278, April
2009. 98
[21] Bethany K Bracken and Gina G Turrigiano. Experience-dependent regulation
of TrkB isoforms in rodent visual cortex. Dev Neurobiol, 69(5):267-278, April
2009. 100
[22] Oliver Braddick and Janette Atkinson. Development of human visual function.
Vision Res, 51(13):1588-1609, July 2011. 95
[23] Nicolas Bray, Harold Pimentel, Pill Melsted, and Lior Pachter. Near-optimal
RNA-Seq quantification. arXiv, May 2015. 49
[24] L Cancedda, E Putignano, and S Impey. Patterned Vision Causes CREMediated Gene Expression in the Visual Cortex through PKA and ERK. The
Journal of ... , 2003. 98
[25] Mehmet Canpolat, Huseyin Per, Hakan Gumus, Ali Yikilmaz, Ekrem Unal,
Turkan Patiroglu, Levent Cinar, Ali Kurtsoy, and Sefer Kumandas. Rapamycin
has a beneficial effect on controlling epilepsy in children with tuberous sclerosis
complex: results of 7 children from a cohort of 86. Childs Nerv Syst, 30(2):227240, June 2013. 56
[26] Valerio Carelli, Fred N Ross-Cisneros, and Alfredo A Sadun. Mitochondrial
dysfunction as a cause of optic neuropathies. Prog Retin Eye Res, 23(1):53-89,
January 2004. 113
[27] G Carmignoto. Activity-dependent decrease in NMDA receptor responses during development of the visual cortex. Science, 1992. 96
[28] M. M. Carrasco, K. A. Razak, and S. L. Pallas. Visual Experience Is Necessary for Maintenance But Not Development of Receptive Fields in Superior
Colliculus. 2005. 95, 116
121
BIBLIOGRAPHY
[29] Maria M Carrasco, Yu-Ting Mao, Timothy S Balmer, and Sarah L Pallas.
Inhibitory plasticity underlies visual deprivation-induced loss of receptive field
refinement in the adult superior colliculus. Eur J Neurosci, 33(1):58-68, November 2010. 92
[30] Maria M Carrasco, Yu-Ting Mao, Timothy S Balmer, and Sarah L Pallas. Inhibitory plasticity underlies visual deprivation-induced loss of receptive field
refinement in the adult superior colliculus. Eur J Neurosci, 33(1):58-68, January 2011. 116
[31] Maria Magdalena Carrasco and Sarah L Pallas. Early visual experience prevents but cannot reverse deprivation-induced loss of refinement in adult superior colliculus. Vis Neurosci, 23(6):845-852, November 2006. 96
[32] Robert P Carson, Dominic L Van Nielen, Peggy A Winzenburger, and Kevin C
Ess. Neuronal and glia abnormalities in Tscl-deficient forebrain and partial
rescue by rapamycin. Neurobiology of Disease, 45(1):369-380, January 2012.
57
[33] V A Casagrande and I T Diamond. Ablation study of the superior colliculus
in the tree shrew (Tupaia glis). J Comp Neurol, 156(2):207-237, July 1974. 88
[34] James Cavanaugh, Ilya E Monosov, Kerry McAlonan, Rebecca Berman,
Mitchell K Smith, Vania Cao, Kuan H Wang, Edward S Boyden, and Robert H
Wurtz. Optogenetic inactivation modifies monkey visuomotor behavior. Neuron, 76(5):901-907, December 2012. 91
[35] Anand R Chandrasekaran, Daniel T Plas, Ernesto Gonzalez, and Michael C
Crair. Evidence for an instructive role of retinal activity in retinotopic map
refinement in the superior colliculus of the mouse. Journal of Neuroscience,
25(29):6929-6938, July 2005. 92
[36] Brad A Chapman. Benchmarking variation and RNA-seq analyses on Amazon
Web Services with Docker. http: //bcb. io/2014/12/19/awsbench/. 18
[37] Brad A Chapman and Rory Kirchner. bcbio-nextgen: automated, validated
analysis of high throughput sequencing data. https: //pypi. python. org/
pypi/bcbio-nextgen/0.9.0, 2015. 12
[38] Dmitri B Chklovskii and Alexei A Koulakov. Maps in the brain: what can we
learn from them? Annu Rev Neurosci, 27:369-392, 2004. 91
122
BIBLIOGRAPHY
[39] Yong-Jin Choi, Alessia Di Nardo, Ioannis Kramvis, Lynsey Meikle, David J
Kwiatkowski, Mustafa Sahin, and Xi He. Tuberous sclerosis complex proteins
control axon formation. Genes & Development, 22(18):2485-2495, September
2008. 55
[40] Francis S Collins and Lawrence A Tabak. Policy: NIH plans to enhance reproducibility. Nature, 505(7485):612-613, January 2014. 23
[41] SEQC MAQC-III Consortium. A comprehensive assessment of RNA-seq accuracy, reproducibility and information content by the Sequencing Quality
Control Consortium. Nat Biotechnol, 32(9):903-914, September 2014. 27
[42] Peter B Crino, Katherine L Nathanson, and Elizabeth Petri Henske. The
Tuberous Sclerosis Complex. N Engl J Med, 355(13):1345-1356, September
2006. 50, 51
[43] Alexis Criscuolo and Sylvain Brisse. AlienTrimmer: a tool to quickly and
accurately trim off multiple short contaminant sequences from high-throughput
sequencing reads. Genomics, 102(5-6):500-506, November 2013. 105
[44] G Del Angel, M A Rivas, M Hanna, and A McKenna. A framework for variation
discovery and genotyping using next-generation DNA sequencing data. Nature
... 2011. 66
[45] David S DeLuca, Joshua Z Levin, Andrey Sivachenko, Timothy Fennell, MarcDanie Nazaire, Chris Williams, Michael Reich, Wendy Winckler, and Gad Getz.
RNA-SeQC: RNA-seq metrics for quality control and process optimization.
Bioinformatics, 28(11):1530-1532, June 2012. 15, 65, 105
[46] Marie-Agnes Dillies, Andrea Rau, Julie Aubert, Christelle Hennequet-Antier,
Marine Jeanmougin, Nicolas Servant, Celine Keime, Guillemette Marot, David
Castel, Jordi Estelle, Gregory Guernec, Bernd Jagla, Luc Jouneau, Denis Lalo8,
Caroline Le Gall, Brigitte Scha~ffer, Stephane Le Crom, Micka8l Guedj, Florence Jaffrezic, and French StatOmique Consortium. A comprehensive evaluation of normalization methods for Illumina high-throughput RNA sequencing
data analysis. Briefings in Bioinformatics, 14(6):671-683, November 2013. 15
[47] Emmanuel Dimont, Jiantao Shi, Rory Kirchner, and Winston Hide. edgeRun:
an R package for sensitive, functionally relevant differential expression discovery using an unconditional exact test. Bioinformatics, page btv209, April 2015.
41
123
BIBLIOGRAPHY
[48] Alexander Dobin, Carrie A Davis, Felix Schlesinger, Jorg Drenkow, Chris Zaleski, Sonali Jha, Philippe Batut, Mark Chaisson, and Thomas R Gingeras.
STAR: ultrafast universal RNA-seq aligner. Bioinformatics, 29(1):15-21, January 2013. 31, 37, 65, 105
[49] R M Douglas, N M Alam, B D Silver, T J McGill, W W Tschetter, and G T
Prusky. Independent visual threshold measurements in the two eyes of freely
moving rats and mice using a virtual-reality optokinetic system. Vis Neurosci,
22(5):677-684, September 2005. 94
[50] Sorin Draghici, Purvesh Khatri, Aron C Eklund, and Zoltan Szallasi. Reliability and reproducibility issues in DNA microarray measurements. Trends
Genet., 22(2):101-109, February 2006. 23
[51] Stephen M Eacker, Matthew J Keuss, Eugene Berezikov, Valina L Dawson, and
Ted M Dawson. Neuronal Activity Regulates Hippocampal miRNA Expression.
PLoS ONE, 6(10):e25068, October 2011. 85
[52] Dan Ehninger, Sangyeul Han, Carrie Shilyansky, Yu Zhou, Weidong Li, David J
Kwiatkowski, Vijaya Ramesh, and Alcino J Silva. Reversal of learning deficits
in a Tsc2+/- mouse model of tuberous sclerosis. Nature medicine, 14(8):843848, August 2008. 55, 56
[53] Dan Ehninger and Alcino J Silva. Rapamycin for treating Tuberous sclerosis
and Autism spectrum disorders. Trends in Molecular Medicine, 17(2):78-87,
February 2011. 56
[54] Par G Engstr6m, Tamara Steijger, Botond Sipos, Gregory R Grant, Andr6
Kahles, Tyler Alioto, Jonas Behr, Paul Bertone, Regina Bohnert, Davide Campagna, Carrie A Davis, Alexander Dobin, Thomas R Gingeras, Nick Goldman, Roderic Guig6, Jennifer Harrow, Tim J Hubbard, G6raldine Jean, Peter Kosarev, Sheng Li, Jinze Liu, Christopher E Mason, Vladimir Molodtsov,
Zemin Ning, Hannes Ponstingl, Jan F Prins, Gunnar Rdtsch, Paolo Ribeca,
Igor Seledtsov, Victor Solovyev, Giorgio Valle, Nicola Vitulo, Kai Wang,
Thomas D Wu, and Georg Zeller. Systematic evaluation of spliced alignment
programs for RNA-seq data. Nat Meth, 10(12):1185-1191, November 2013. 15,
37
[55] B Ewing, L Hillier, M C Wendl, and P Green. Base-calling of automated
sequencer traces using phred. I. Accuracy assessment. Genome Res, 8(3):175185, March 1998. 105
124
BIBLIOGRAPHY
125
-
[56] M Fagiolini, T Pizzorusso, N Berardi, and L Domenici. ScienceDirect.com
Vision Research - Functional postnatal development of the rat primary visual
cortex and the role of visual experience: Dark rearing and monocular deprivation. Vision Res, 1994. 95
[57] Michela Fagiolini, Hiroyuki Katagiri, Hiroyuki Miyamoto, Hisashi Mori, Seth
G N Grant, Masayoshi Mishina, and Takao K Hensch. Separable features of visual cortical plasticity revealed by N-methyl-D-aspartate receptor 2A signaling.
Proc Natl Acad Sci USA, 100(5):2854-2859, March 2003. 95
[58] N Fillmore, Y Bai, M Collins, J A Thomson, and R Stewart. Evaluation of de
novo transcriptome assemblies from RNA-Seq data. bioRxiv, 2014. 15
[59] I Fine, AR Wade, AA Brewer, and MG May. Long-term deprivation affects
visual perception and cortex - Nature Neuroscience. Neuroscience, 2003. 95
[60] Alyssa C Frazee, Sarven Sabunciyan, Kasper D Hansen, Rafael A Irizarry, and
Jeffrey T Leek. Differential expression analysis of RNA-seq data at single-base
resolution. Biostatistics (Oxford, England), 15(3):413-426, July 2014. 40
[61] T W Frazier, R Embacher, A K Tilot, K Koenig, J Mester, and C Eng. Molecular and phenotypic abnormalities in individuals with germline heterozygous
PTEN mutations and autism. Mol Psychiatry, October 2014. 60
[62] Fernando Garcia-Alcalde, Konstantin Okonechnikov, Jose Carbonell, Luis M
Cruz, Stefan G6tz, Sonia Tarazona, Joaquin Dopazo, Thomas F Meyer, and
Ana Conesa. Qualimap: evaluating next-generation sequencing alignment data.
Bioinformatics, 28(20):2678-2679, October 2012. 15
[63] Shruti Garg, Jonathan Green, Kathy Leadbitter, Richard Emsley, Annukka
Lehtonen, D Gareth Evans, and Susan M Huson. Neurofibromatosis type 1
and autism spectrum disorder. Pediatrics, 132(6):e1642-8, December 2013. 60
[64] Julie Goldberg. Eye-opening dependent elaboration and refinement of the cortical projection to the superior colliculus in rats. Massachusetts Institute of
Technology, pages 1-89, August 2009. 98, 99, 106, 107, 108, 116
[65] Michael E Goldberg and Robert H Wurtz. Activity of superior colliculus in
behaving monkey. I. Visual receptive fields of single neurons. J Neurophysiol,
35(4):542-559, 1972. 89
BIBLIOGRAPHY
[66] Elena A Goncharova, Dmitry A Goncharov, Andrew Eszterhas, Deborah S
Hunter, Marilyn K Glassberg, Raymond S Yeung, Cheryl L Walker, Daniel
Noonan, David J Kwiatkowski, Margaret M Chou, Reynold A Panettieri, and
Vera P Krymskaya. Tuberin regulates p70 S6 kinase activation and ribosomal
protein S6 phosphorylation. A role for the TSC2 tumor suppressor gene in
pulmonary lymphangioleiomyomatosis (LAM). J Biol Chem, 277(34):3095830967, August 2002. 53
[67] T Griebel, B Zacher, P Ribeca, E Raineri, V Lacroix, R Guigo, and M Sammeth. Modelling and simulating generic RNA-Seq experiments with the flux
simulator. Nucleic Acids Res, 40(20):10073-10083, November 2012. 28
[68] Dominic Grin, Lennart Kester, and Alexander van Oudenaarden. Validation
of noise models for single-cell transcriptomics. Nat Meth, 11(6):637-640, June
2014. 48
[69] Shuchen Gu, Michalis Kounenidakis, Eva-Maria Schmidt, Divija Deshpande, Saad Alkahtani, Saud Alarifi, Michael F61ler, Konstantinos Alevizopoulos, Florian Lang, and Christos Stournaras.
Rapid activation of
FAK/mTOR/p70S6K/PAK 1-signaling controls the early testosterone-induced
actin reorganization in colon cancer cells. CellularSignalling, 25(1):66-73, January 2013. 60
[70] Simone Gupta, Shannon E Ellis, Foram N Ashar, Anna Moes, Joel S Bader,
Jianan Zhan, Andrew B West, and Dan E Arking. Transcriptome analysis
reveals dysregulation of innate immune response genes and neuronal activitydependent genes in autism. Nature Communications, 5:5748, 2014. 62
[71] William C Hall and Adonis K Moschovakis, editors. The Superior Colliculus:
New Approaches for Studying Sensorimotor Integration (Methods and New
Frontiers in Neuroscience). CRC Press, 1 edition, September 2003. 87
[72] Thomas J Hardcastle and Krystyna A Kelly. baySeq: Empirical Bayesian
methods for identifying differential expression in sequence count data. BMC
Bioinformatics, 11(1):422, August 2010. 41
[73] J K Harting, I T Diamond, and W C Hall. Anterograde degeneration study of
the cortical projections of the lateral geniculate and pulvinar nuclei in the tree
shrew (Tupaia glis). J Comp Neurol, 150(4):393-439, 2004. 88
126
BIBLIOGRAPHY
[74] D 0 Hebb. The Organization of Behavior. A Neuropsychological Theory.
Psychology Press, January 2002. 93
[75] G F L Hofbauer, A Marcollo-Pini, A Corsenca, A D Kistler, L E French,
R P Wfithrich, and A L Serra. The mTOR inhibitor rapamycin significantly
improves facial angiofibroma lesions in a patient with tuberous sclerosis. Br.
J. Dermatol., 159(2):473-475, August 2008. 53
[76] G HUDSON, S KEERS, P MAN, P GRIFFITHS, K HUOPONEN, M SAVONTAUS, E NIKOSKELAINEN, M ZEVIANI, F CARRARA, and R HORVATH.
Identification of an X-Chromosomal Locus and Haplotype Modulating the Phenotype of a Mitochondrial DNA Disorder. The American Journal of Human
Genetics, 77(6):1086-1091, December 2005. 113
[77] Darrel C Ince, Leslie Hatton, and John Graham-Cumming. The case for open
computer programs. Nature, 482(7386):485-488, February 2012. 23
[78] Broad Institute. Calling variants in RNAseq. https: //www. broadinst itute.
org/gatk/guide/article?id=3891, 2014. 45, 66
[79] Manuel Irimia, Robert J Weatheritt, Jonathan D Ellis, Neelroop N Parikshak,
Thomas Gonatopoulos-Pournatzis, Mariana Babor, Mathieu Quesnel-Vallieres,
Javier Tapial, Bushra Raj, Dave O'Hanlon, Miriam Barrios-Rodiles, Michael
J E Sternberg, Sabine P Cordes, Frederick P Roth, Jeffrey L Wrana, Daniel H
Geschwind, and Benjamin J Blencowe. A highly conserved program of neuronal microexons is misregulated in autistic brains. Cell, 159(7):1511-1523,
December 2014. 48
[80] T Isa, T Endo, and Y Saito. The visuo-motor pathway in the local circuit of
the rat superior colliculus. J Neurosci, 18(20):8496-8504, October 1998. 89
[81] Saiful Islam, Amit Zeisel, Simon Joost, Gioele La Manno, Pawel Zajac, Maria
Kasper, Peter L6nnerberg, and Sten Linnarsson. Quantitative single-cell RNAseq with unique molecular identifiers. Nat Meth, 11(2):163-166, February 2014.
48
[82] Yanli Ji, Xiaoyun Jia, Shiqiang Li, Xueshan Xiao, Xiangming Guo, and
Qingjiong Zhang. Evaluation of the X-linked modifier loci for Leber hereditary optic neuropathy with the G11778A mutation in Chinese. Mol. Vis.,
16:416-424, 2010. 113
127
BIBLIOGRAPHY
[83] Hongshan Jiang, Rong Lei, Shou-Wei Ding, and Shuifang Zhu. Skewer: a fast
and accurate adapter trimmer for next-generation sequencing paired-end reads.
BMC Bioinformatics, 15(1):182, 2014. 35
[84] C JOINSON, F J O'CALLAGHAN, J P OSBORNE, C MARTYN, T HARRIS,
and P F BOLTON. Learning disability and epilepsy in an epidemiological
sample of individuals with tuberous sclerosis complex. Psychological Medicine,
33(02):335-344, February 2003. 50
[85] James A Kaltenbach and John Lazor. Tonotopic maps obtained from the
surface of the dorsal cochlear nucleus of the hamster and rat. Hearing research,
51(1):149-160, January 1991. 91
[86] M Kanehisa and S Goto. KEGG: kyoto encyclopedia of genes and genomes.
Nucleic acids research, 2000. 75
[87] Y Katz, E Wang, and E Airoldi. Analysis and design of RNA sequencing
experiments for identifying isoform regulation. Nat Meth, 2010. 40
[88] Do-Hyung Kim, Dos D Sarbassov, Siraj M Ali, Jessie E King, Robert R Latek,
Hediye Erdjument-Bromage, Paul Tempst, and David M Sabatini. mTOR
Interacts with Raptor to Form a Nutrient-Sensitive Complex that Signals to
the Cell Growth Machinery. Cell, 110(2):163-175, July 2002. 53
[89] Stefan Kirov, Ruiru Ji, Jing Wang, and Bing Zhang. Functional annotation
of differentially regulated gene set using WebGestalt: a gene set predictive of
response to ipilimumab in tumor biopsies. Methods Mol Biol, 1101(Chapter
3):31-42, 2014. 66
[90] Paul J Kiser, Zijing Liu, Steven D Wilt, and George D Mower. Cellular and
laminar expression of Dab-1 during the postnatal critical period in cat visual
cortex and the effects of dark rearing. Brain Res, 1383:81-89, April 2011. 98
[91] T Kobayashi, 0 Minowa, Y Sugitani, S Takai, H Mitani, E Kobayashi, T Noda,
and 0 Hino. A germ-line Tscl mutation causes tumor development and embryonic lethality that are similar, but not identical to, those caused by Tsc2
mutation in mice. Proc Natl Acad Sci USA, 98(15):8762-8767, July 2001. 56
[92] Sek Won Kong, Mustafa Sahin, Christin D Collins, Mary H Wertz, Malcolm G
Campbell, Jarrett D Leech, Dilja Krueger, Mark F Bear, Louis M Kunkel, and
Isaac S Kohane. Divergent dysregulation of gene expression in murine models
of fragile X syndrome and tuberous sclerosis. Mol Autism, 5(1):16, 2014. 62
128
BIBLIOGRAPHY
[93] K Krug, C J Akerman, and I D Thompson. Responses of neurons in neonatal
cortex and thalamus to patterned visual stimulation through the naturally
closed lids. J Neurophysiol, 85(4):1436-1443, April 2001. 94
[94] C W Law, Y Chen, W Shi, and G K Smyth. Voom: precision weights unlock
linear model analysis tools for RNA-seq read counts. Preprint 2013, 2013. 28,
32, 41
[95] Ning Leng, John A Dawson, James A Thomson, Victor Ruotti, Anna I Rissman,
Bart M G Smits, Jill D Haag, Michael N Gould, Ron M Stewart, and Christina
Kendziorski. EBSeq: an empirical Bayes hierarchical model for inference in
RNA-seq experiments. Bioinformatics, 29(8):1035-1043, April 2013. 41
[96] Nanxin Li, Boyoung Lee, Rong-Jian Liu, Mounira Banasr, Jason M Dwyer,
Masaaki Iwata, Xiao-Yuan Li, George Aghajanian, and Ronald S Duman.
mTOR-dependent synapse formation underlies the rapid antidepressant effects
of NMDA antagonists. Science, 329(5994):959-964, August 2010. 55
[97] Yang Liao, Gordon K Smyth, and Wei Shi. featureCounts: an efficient general
purpose program for assigning sequence reads to genomic features. Bioinformatics, 30(7):923-930, April 2014. 40, 66, 105
[98] Robert Lindner and Caroline C Friedel. A Comprehensive Evaluation of Alignment Algorithms in the Context of RNA-Seq. PLoS ONE, 7(12):e52403, 2012.
26
[99] Jonathan 0 Lipton and Mustafa Sahin. The Neurology of mTOR. Neuron,
84(2):275-291, October 2014. 52
[100] S G Lomber, B R Payne, and P Cornwell. Role of the superior colliculus
in analyses of space: superficial and intermediate layer contributions to visual
orienting, auditory orienting, and visuospatial discriminations during unilateral
and bilateral deactivations. J Comp Neurol, 441(1):44-57, December 2001. 88
[101] M I Love, W Huber, and S Anders. Moderated estimation of fold change and
dispersion for RNA-Seq data with DESeq2. Genome Biol, 2014. 32, 41
[102] Joaquin N Lugo, Gregory D Smith, Erin P Arbuckle, Jessika White, Andrew J
Holley, Crina M Floruta, Nowrin Ahmed, Maribel C Gomez, and Obi Okonkwo.
Deletion of PTEN produces autism-like behavioral deficits and alterations in
synaptic proteins. Front Mol Neurosci, 7:27, 2014. 60
129
BIBLIOGRAPHY
130
[103] R D Lund. Terminal Distribution in the Superior Colliculus of Fibres Originating in the Visual Cortex. Nature, 204(4965):1283-1285, December 1964.
89
[104] Daniel Macarthur. Methods: Face up to false positives. Nature, 487(7408):427428, July 2012. 23
[105] Matthew D Macmanes. On the optimal trimming of high-throughput mRNA
sequence data. Front Genet, 5:13, 2014. 35, 65, 105
[106] Marta Majdan and Carla J Shatz. Effects of visual experience on activitydependent gene regulation in cortex. Nat Neurosci, 9(5):650-659, May 2006.
97, 116
[107] A Mansour, J H Meador-Woodruff, Q Zhou, 0 Civelli, H Akil, and S J Watson. A comparison of D1 receptor binding and mRNA in rat brain using
receptor autoradiographic and in situ hybridization techniques. Neuroscience,
46(4):959-971, February 1992. 114
[108] A Mansour and S J Watson.
Dopamine receptor expression in the central
nervous system. Psychopharmacology: the fourth generation of ... , 1995. 114
[109] Marcel Martin. Cutadapt removes adapter sequences from high-throughput
sequencing reads. EMBnet.journal, 17(l):pp. 10-12, February 2011. 35, 65
[110] Daphne Maurer, Terri L Lewis, and Catherine J Mondloch. Missing sights:
consequences for visual cognitive development. Trends Cogn. Sci. (Regul. Ed.),
9(3):144-151, March 2005. 95
[111] Christopher J McDougle, Lawrence Scahill, Michael G Aman, James T McCracken, Elaine Tierney, Mark Davies, L Eugene Arnold, David J Posey, Andres Martin, Jaswinder K Ghuman, Bhavik Shah, Shirley Z Chuang, Naomi B
Swiezy, Nilda M Gonzalez, Jill Hollway, Kathleen Koenig, James J McGough,
Louise Ritz, and Benedetto Vitiello. Risperidone for the Core Symptom Domains of Autism: Results From the Study by the Autism Network of the
Research Units on Pediatric Psychopharmacology. American Journal of Psychiatry, 162(6):1142-1148, November 2014. 86
[112] Todd McLaughlin, Christine L Torborg, Marla B Feller, and Dennis D M
O'Leary. Retinotopic map refinement requires spontaneous retinal waves during
a brief critical period of development. Neuron, 40(6):1147-1160, December
2003. 92, 93
BIBLIOGRAPHY
[113] John J McMahon, Wilson Yu, Jun Yang, Haihua Feng, Meghan Helm, Elizabeth McMahon, Xinjun Zhu, Damian Shin, and Yunfei Huang. Seizuredependent mTOR activation in 5-HT neurons promotes autism-like behaviors
in mice. Neurobiology of Disease, 73:296-306, January 2015. 81
[114] Lynsey Meikle, Delia M Talos, Hiroaki Onda, Kristen Pollizzi, Alexander
Rotenberg, Mustafa Sahin, Frances E Jensen, and David J Kwiatkowski. A
mouse model of tuberous sclerosis: neuronal loss of Tscl causes dysplastic and
ectopic neurons, reduced myelination, seizure activity, and limited survival.
Journal of Neuroscience, 27(21):5546-5558, May 2007. 55, 57, 63, 64
[115] Nikolaos Mellios, Hiroki Sugihara, Jorge Castro, Abhishek Banerjee, Chuong
Le, Arooshi Kumar, Benjamin Crawford, Julia Strathmann, Daniela Tropea,
Stuart S Levine, Dieter Edbauer, and Mriganka Sur. miR-132, an experiencedependent microRNA, is essential for visual cortex plasticity. Nat Neurosci,
14(10):1240-1242, October 2011. 98
[116] M M Merzenich, J H Kaas, J Wall, R J Nelson, M Sur, and D Felleman. Topographic reorganization of somatosensory cortical areas 3b and 1 in adult
monkeys following restricted deafferentation. Neuroscience, 8(1):33-55, January 1983. 91
[117] Masashi Mizuguchi and Sachio Takashima. Neuropathology of tuberous sclerosis. Brain Dev, 23(7):508-515, November 2001. 51
[118] Ahmad R Mohamed, Catherine A Bailey, Jeremy L Freeman, Wirginia
Maixner, Graeme D Jackson, and A Simon Harvey. Intrinsic epileptogenicity of cortical tubers revealed by intracranial EEG monitoring. Neurology,
79(23):2249-2257, December 2012. 52
[119] Andrei I Molosh, Philip L Johnson, John P Spence, David Arendt, Lauren M
Federici, Cristian Bernabe, Steven P Janasik, Zaneer M Segu, Rajesh Khanna,
Chirayu Goswami, Weiguo Zhu, Su-Jung Park, Lang Li, Yehia S Mechref,
D Wade Clapp, and Anantha Shekhar. Social learning and amygdala disruptions in Nfl mice are rescued by blocking p21-activated kinase. Nat Neurosci,
17(11):1583-1590, November 2014. 60
[120] S Molotchnikoff and S K Itaya. Functional development of the neonatal rat
retinotectal pathway. Brain Res. Dev. Brain Res., 72(2):300-304, April 1993.
93
131
132
BIBLIOGRAPHY
[121] Ali Mortazavi, Brian A Williams, Kenneth Mccue, Lorian Schaeffer, and Barbara Wold. Mapping and quantifying mammalian transcriptomes by RNA-Seq.
Nat Meth, 5(7):621, May 2008. 100
[122] G D Mower, C J Caplan, and G Letsou. Behavioral recovery from binocular
deprivation in the cat. Behav Brain Res, 4(2):209-215, February 1982. 95
[123] Elly Nedivi. Molecular analysis of developmental plasticity in neocortex. J.
Neurobiol., 41(1):135-147, October 1999. 96, 97
[124] Anton Nekrutenko and James Taylor. Next-generation sequencing data interpretation: enhancing reproducibility and accessibility. Nat Rev Genet,
13(9):667-672, September 2012. 23
[125] Stanley F Nelson. [Book Review: Microarray Gene Expression Data Analysis: A Beginner's Guide]. The Quarterly Review of Biology, 79(3):308-309,
September 2004. 12
[126] Lena H Nguyen, Amy L Brewster, Madeline E Clark, Angelique RegnierGolanov, C Nicole Sunnen, Vinit V Patil, Gabriella D'Arcangelo, and Anne E
Anderson. mTOR inhibition suppresses established epilepsy in a mouse model
of cortical dysplasia. Epilepsia, March 2015. 60
[127] Duyu Nie, Alessia Di Nardo, Juliette M Han, Hasani Baharanyi, loannis
Kramvis, ThanhThao Huynh, Sandra Dabora, Simone Codeluppi, Pier Paolo
Pandolfi, Elena B Pasquale, and Mustafa Sahin. Tsc2-Rheb signaling regulates
EphA-mediated axon guidance. Nat Neurosci, 13(2):163-172, January 2010.
55
[128] Umadevi Paila, Brad A Chapman, Rory Kirchner, and Aaron R Quinlan. GEMINI: integrative exploration of genetic variation and genome annotations. PLoS
Comput. Biol., 9(7):e1003153, 2013. 14, 18, 21, 46
[129] L Pantano, X Estivill, and E Marti. SeqBuster, a bioinformatic tool for the
processing and analysis of small RNAs datasets, reveals ubiquitous miRNA
modifications in human embryonic cells. Nucleic acids research, 2010. 14
[130] Rob Patro.
Salmon.
salmon.html, 2015. 49
http://sailfish.readthedocs.org/en/latest/
BIBLIOGRAPHY
[131] M Paulussen, L Van Brussel, and L Arckens. Monocular enucleation profoundly
reduces secretogranin II expression in adult mouse visual cortex. Neurochem.
Int., 59(7):1082-1094, December 2011. 98
[132] H Peng, J Liu, Q Sun, R Chen, Y Wang, J Duan, C Li, B Li, Y Jing, X Chen,
Q Mao, K-F Xu, C L Walker, J Li, J. Wang, and H Zhang. mTORC1 enhancement of STIM1-mediated store-operated Ca2+ entry constrains tuberous sclerosis complex-related tumor development. Oncogene, 32(39):4702-4711,
September 2013. 81
[133] Fernando Perez and Brian E Granger. IPython: A System for Interactive
Scientific Computing. Computing in Science &amp; Engineering, 9(3):21-29,
2007. 18
[134] Geo Pertea. high FPKM called with few reads mapping to transcript - Issue
#12 - gpertea/stringtie. https ://github. com/gpertea/stringtie/issues/
12. 49
[135] Mihaela Pertea, Geo M Pertea, Corina M Antonescu, Tsung-Cheng Chang,
Joshua T Mendell, and Steven L Salzberg. StringTie enables improved reconstruction of a transcriptome from RNA-seq reads. Nat Biotechnol, February
2015. 49
[136] Marnie A Phillips. Eye-opening and the control of visual synapse development
in the mouse superior colliculus. Massachusetts Institute of Technology, pages
1-128, September 2007. 90
[137] Marnie A Phillips, Matthew T Colonnese, Julie Goldberg, Laura D Lewis,
Emery N Brown, and Martha Constantine-Paton. A synaptic strategy for
consolidation of convergent visuotopic maps. Neuron, 71(4):710-724, August
2011. 98, 116
[138] Nathalie Picard, Jennifer H Leslie, Sara K Trowbridge, Jaichandar Subramanian, Elly Nedivi, and Michela Fagiolini. Aberrant development and plasticity
of excitatory visual cortical networks in the absence of cpgl5. Journal of Neuroscience, 34(10):3517-3522, March 2014. 116
[139] Joseph K Pickrell, Yoav Gilad, and Jonathan K Pritchard. Comment on
"Widespread RNA and DNA sequence differences in the human transcriptome".
Science, 335(6074):1302-author reply 1302, March 2012. 45
133
BIBLIOGRAPHY
[140] Ellen Plasschaert, Mie-Jef Descheemaeker, Lien Van Eylen, Ilse Noens, Jean
Steyaert, and Eric Legius. Prevalence of Autism Spectrum Disorder symptoms
in children with neurofibromatosis type 1. Am. J. Med. Genet. B Neuropsychiatr. Genet., 168B(1):72-80, January 2015. 60
[141] J Roxanne Prichard, Hilda S Armacanqui, Ruth M Benca, and Mary Behan.
Light-dependent retinal innervation of the rat superior colliculus. Anat Rec
(Hoboken), 290(3):341-348, March 2007. 116
[142] JR Prichard and HS Armacanqui. Light-dependent retinal innervation of the
rat superior colliculus - Prichard - 2007 - The Anatomical Record: Advances
in Integrative Anatomy and Evolutionary Biology - Wiley Online Library.
Record: Advances in . . , 2007. 95
[143] C Quaia, H Aizawa, L M Optican, and R H Wurtz. Reversible inactivation
of monkey superior colliculus. II. Maps of saccadic deficits. J Neurophysiol,
79(4):2097-2110, April 1998. 90
[144] Franck Rapaport, Raya Khanin, Yupu Liang, Mono Pirun, Azra Krek, Paul
Zumbo, Christopher E Mason, Nicholas D Socci, and Doron Betel. Comprehensive evaluation of differential gene expression analysis methods for RNA-seq
data. Genome Biol, 14(9):R95, 2013. 15, 26
[145] Khaleel A Razak and Sarah L Pallas. Dark rearing reveals the mechanism
underlying stimulus size tuning of superior colliculus neurons. Vis Neurosci,
23(5):741-748, September 2006. 116
[146] Adam Roberts and Lior Pachter. Streaming fragment assignment for real-time
analysis of sequencing experiments. Nat Meth, 10(1):71-73, January 2013. 37,
40
[147] Adam Roberts and Lior Pachter. Streaming fragment assignment for real-time
analysis of sequencing experiments. Nat Meth, 10(1):71-73, January 2013. 66
[148] Adam Roberts, Harold Pimentel, Cole Trapnell, and Lior Pachter. Identification of novel transcripts in annotated genomes using RNA-Seq. Bioinformatics,
27(17):2325-2329, September 2011. 42
[149] Mark D Robinson, Davis J McCarthy, and Gordon K Smyth. edgeR: a Bioconductor package for differential expression analysis of digital gene expression
data. Bioinformatics, 26(1):139-140, January 2010. 32, 41, 108, 110
134
BIBLIOGRAPHY
[150] Natalia S Rozas, John B Redell, James McKenna, Anthony N Moore, Michael J
Gambello, and Pramod K Dash. Prolonging the survival of Tsc2 conditional
knockout mice by glutamine supplementation. Biochem. Biophys. Res. Commun., January 2015. 56
[151] Atsushi Sato, Shinya Kasai, Toshiyuki Kobayashi, Yukio Takamatsu, Okio
Hino, Kazutaka Ikeda, and Masashi Mizuguchi. Rapamycin reverses impaired
social interaction in mouse models of tuberous sclerosis complex. Nature Communications, 3:1292, December 2012. 56
[152] Yiannis A Savva, Leila E Rieder, and Robert A Reenan. The ADAR protein
family. Genome Biol, 13(12):252, 2012. 46
[153] P H Schiller, J H Sandell, and J H Maunsell. The effect of frontal eye field
and superior colliculus lesions on saccadic latencies in the rhesus monkey. J
Neurophysiol, 57(4):1033-1049, April 1987. 88
[154] Mark Seger. collectl. http: //collectl. sourcef orge. net, 2015. 18
[155] Suma P Shankar, John H Fingert, Valerio Carelli, Maria L Valentino,
Terri M King, Stephen P Daiger, Solange R Salomao, Adriana Berezovsky,
Jr Rubens Belfort, Terri A Braun, Val C Sheffield, Alfredo A Sadun, and
Edwin M Stone. Evidence for a Novel X-Linked Modifier Locus for Leber
Hereditary Optic Neuropathy. http://dx.doi.org/10.1080/13816810701867607,
29(1):17-24, July 2009. 113
[156] Ali Sharma, Charles A Hoeffer, Yukihiro Takayasu, Takahiro Miyawaki,
Sean M McBride, Eric Klann, and R Suzanne Zukin. Dysregulation of mTOR
Signaling in Fragile X Syndrome. J Neurosci, 30(2):694-702, January 2010. 60
[157] Jill L Silverman, Mu Yang, Catherine Lord, and Jacqueline N Crawley. Behavioural phenotyping assays for mouse models of autism. Nat Rev Neurosci,
11(7):490-502, July 2010. 59
[158] T Soucek, G H61zl, G Bernaschek, and M Hengstschlager. A role of the
tuberous sclerosis gene-2 product during neuronal differentiation. Oncogene,
16(17):2197-2204, April 1998. 53
[159] Thomas Soucek, Oliver Pusch, Ralf Wienecke, Jeffrey E DeClue, and Markus
Hengstschldger. Role of the Tuberous Sclerosis Gene-2 Product in Cell Cycle
Control: loss of the tuberous sclerosis gene-2 induces quiescent cells to enter s
phase. J Biol Chem, 272(46):29301-29308, November 1997. 53
135
BIBLI0GRAPHY
136
[160] D.L. Sparks and R Hartwich-Young. The deep layers of the superior colliculus.
Reviews of oculomotor research, 3:213, 1989. 89
[161] R W Sperry. Chemoaffinity in the orderly growth of nerve fiber patterns and
connections. Proc Natl Acad Sci USA, 50(4):703, October 1963. 91
[162] CWL standards team. Common Workflow Language, Draft 2.
common-workf low-language. github. io/#/, 2015. 49
http://
[163] Tamara Steijger, Josep F Abril, Pir G Engstr6m, Felix Kokocinski, Tim J
Hubbard, Roderic Guig6, Jennifer Harrow, Paul Bertone, and RGASP Consortium. Assessment of transcript reconstruction methods for RNA-seq. Nat
Meth, 10(12):1177-1184, December 2013. 15, 42
[164] W Richard Stevens and Stephen A Rago. Advanced Programming in the UNIX
Environment. Addison-Wesley, June 2013. 35
[1651 M P Stryker and W A Harris. Binocular impulse blockade prevents the formation of ocular dominance columns in cat visual cortex. J Neurosci, 6(8):21172133, August 1986. 93
[166] Guomei Tang, Kathryn Gudsnuk, Sheng-Han Kuo, Marisa L Cotrina, Gorazd
Rosoklija, Alexander Sosunov, Mark S Sonders, Ellen Kanter, Candace
Castagna, Ai Yamamoto, Zhenyu Yue, Ottavio Arancio, Bradley S Peterson,
Frances Champagne, Andrew J Dwork, James Goldman, and David Sulzer.
Loss of mTOR-dependent macroautophagy causes autistic-like synaptic pruning deficits. Neuron, 83(5):1131-1143, September 2014. 59
[167] Paola Tognini, Elena Putignano, Alessandro Coatti, and Tommaso Pizzorusso.
Experience-dependent expression of miR-132 regulates ocular dominance plasticity. Nat Neurosci, 14(10):1237-1239, October 2011. 98
[168] M Townsend. Retina-driven dephosphorylation of the NR2A subunit correlates
with faster NMDA receptor kinetics at developing retinocollicular synapses.
Journal of Neuroscience, 24(49):11098-11107, December 2004. 96
[169] M Townsend, A Yoshii, M Mishina, and M Constantine-Paton. Developmental
loss of miniature N-methyl-D-aspartate receptor currents in NR2A knockout
mice. Proc Natl Acad Sci USA, 100(3):1340-1345, February 2003. 97, 116
[170] Col Trapnell, Lio Pachter, and Steven Salzberg. TopHat: discovering splice
junctions with RNA-Seq. Bioinformatics, 25(9):0, May 2009. 33
BIBLIOGRAPHY
137
[171] Cole Trapnell, Adam Roberts, Loyal Goff, Geo Pertea, Daehwan Kim, David R
Kelley, Harold Pimentel, Steven L Salzberg, John L Rinn, and Lior Pachter.
Differential gene and transcript expression analysis of RNA-seq experiments
with TopHat and Cufflinks. Nat Protoc, 7(3):562-578, March 2012. 41
[172] Cole Trapnell, Brian A Williams, Geo Pertea, Ali Mortazavi, Gordon Kwan,
Marijke J van Baren, Steven L Salzberg, Barbara J Wold, and Lior Pachter.
Transcript assembly and quantification by RNA-Seq reveals unannotated transcripts and isoform switching during cell differentiation.
Nat Biotechnol,
28(5):511, May 2010. 40
[173] Jason W Triplett, An Phan, Jena Yamada, and David A Feldheim. Alignment
of Multimodal Sensory Input in the Superior Colliculus through a GradientMatching Mechanism. Journal of Neuroscience, 32(15):5264-5271, April 2012.
92
[174] Daniela Tropea, Gabriel Kreiman, Alvin Lyckman, Sayan Mukherjee, Hongbo
Yu, Sam Horng, and Mriganka Sur. Gene expression changes and molecular
pathways mediating activity-dependent plasticity in visual cortex. Nat Neurosci, 9(5):660-668, May 2006. 97, 98, 116
[175] Daniela Tropea, Audra Van Wart, and Mriganka Sur. Molecular mechanisms
of experience-dependent plasticity in visual cortex. Philos Trans R Soc Lond,
B, Biol Sci, 364(1515):341-355, February 2009. 97
[176] R J Tusa, M J Mustari, A F Burrows, and A F Fuchs. Gaze-stabilizing deficits
and latent nystagmus in monkeys with brief, early-onset visual deprivation:
eye movement recordings. J Neurophysiol, 86(2):651-661, August 2001. 95
[177] Erik J Uhlmann, Michael Wong, Rebecca L Baldwin, M Livia Bajenaru, Hiroaki Onda, David J Kwiatkowski, Kelvin Yamada, and David H Gutmann.
Astrocyte-specific TSC1 conditional knockout mice exhibit abnormal neuronal
organization and seizures. Ann Neurol., 52(3):285-296, September 2002. 57
[178] Geraldine
Van
der
Auwera.
Calling
variants
on
cohorts
of
samples
using
the
HaplotypeCaller
in
GVCF
mode.
http: //gatkforums . broadinstitute. org/discussion/3893/
calling-variants-on-cohorts-of -samples-using-the-haplotype caller-in-gvcf -mode,
2014. 67
BIBLIOGRAPHY
[179] Jeremy Veenstra-VanderWeele, Christopher L Muller, Hideki Iwamoto, Jennifer E Sauer, W Anthony Owens, Charisma R Shah, Jordan Cohen, Padmanabhan Mannangatti, Tammy Jessen, Brent J Thompson, Ran Ye, Travis M
Kerr, Ana M Carneiro, Jacqueline N Crawley, Elaine Sanders-Bush, Douglas G
McMahon, Sammanda Ramamoorthy, Lynette C Daws, James S Sutcliffe, and
Randy D Blakely. Autism gene variant causes hyperserotonemia, serotonin receptor hypersensitivity, social impairment and repetitive behavior. Proc Natl
Acad Sci USA, 109(14):5469-5474, April 2012. 81
[180] Irina Voineagu, Xinchen Wang, Patrick Johnston, Jennifer K Lowe, Yuan Tian,
Steve Horvath, Jonathan Mill, Rita M Cantor, Benjamin J Blencowe, and
Daniel H Geschwind. Transcriptomic analysis of autistic brain reveals convergent molecular pathology. Nature, 474(7351):380-384, June 2011. 62, 63
[181] Eric T Wang, Rickard Sandberg, Shujun Luo, Irina Khrebtukova, Lu Zhang,
Christine Mayr, Stephen F Kingsmore, Gary P Schroth, and Christopher B
Burge. Alternative isoform regulation in human tissue transcriptomes. Nature,
456(7221):470-476, November 2008. 38
[182] L Wang, R Sarnaik, and K Rangarajan. Visual Receptive Field Properties of
Neurons in the Superficial Superior Colliculus of the Mouse. The Journal of
2010. 95
[183] Liguo Wang, Hyun Jung Park, Surendra Dasari, Shengqin Wang, Jean-Pierre
Kocher, and Wei Li. CPAT: Coding-Potential Assessment Tool using an
alignment-free logistic regression model. Nucleic Acids Res, 41(6):e74-e74,
April 2013. 42
[184] Lupeng Wang, Krsna V Rangarajan, Courtney A Lawhn-Heath, Rashmi Sarnaik, Bor-Shuen Wang, Xiaorong Liu, and Jianhua Cang. Direction-specific
disruption of subcortical visual behavior and receptive fields in mice lacking
the beta2 subunit of nicotinic acetylcholine receptor. Journal of Neuroscience,
29(41):12909-12918, October 2009. 93
[185] Yanling Wang, Joel S F Greenwood, Maria Elisa Calcagnotto, Heidi E Kirsch,
Nicholas M Barbaro, and Scott C Baraban. Neocortical hyperexcitability in a
human case of tuberous sclerosis complex and mice lacking neuronal expression
of TSC1. Ann Neurol., 61(2):139-152, February 2007. 52, 58
[186] Zhong Wang, Mark Gerstein, and Michael Snyder. RNA-Seq: a revolutionary
tool for transcriptomics. Nat Rev Genet, 10(1):57-63, 2009. 13
138
BIBLIOGRAPHY
[187] D M Weiner, A I Levey, R K Sunahara, H B Niznik, B F O'Dowd, P Seeman,
and M R Brann. D1 and D2 dopamine receptor mRNA in rat brain. Proc Natl
Acad Sci USA, 88(5):1859-1863, March 1991. 114
[188] R.O.L. Wong. Retinal waves and visual system development. Annu Rev Neurosci, 22(1):29-47, 1999. 92, 93
[189] Margaret T T Wong-Riley. Energy metabolism of the visual system. Eye Brain,
2:99-116, 2010. 113
[190] Michael Yourshaw, S Paige Taylor, Aliz R Rao, Martin G Martin, and Stanley F Nelson. Rich annotation of DNA sequencing variants by leveraging the
Ensembl Variant Effect Predictor with plugins. Briefings in Bioinformatics,
page bbu008, March 2014. 67
[191] Ryan K C Yuen, Bhooma Thiruvahindrapuram, Daniele Merico, Susan
Walker, Kristiina Tammimies, Ny Hoang, Christina Chrysler, Thomas Nalpathamkalam, Giovanna Pellecchia, Yi Liu, Matthew J Gazzellone, Lia
D'Abate, Eric Deneault, Jennifer L Howe, Richard S C Liu, Ann Thompson,
Mehdi Zarrei, Mohammed Uddin, Christian R Marshall, Robert H Ring, Lonnie Zwaigenbaum, Peter N Ray, Rosanna Weksberg, Melissa T Carter, Bridget A Fernandez, Wendy Roberts, Peter Szatmari, and Stephen W Scherer.
Whole-genome sequencing of quartet families with autism spectrum disorder.
Nature medicine, 21(2):185-191, January 2015. 59
[192] Ling-Hui Zeng, Yannan Ouyang, Vered Gazit, John R Cirrito, Laura A Jansen,
Kevin C Ess, Kelvin A Yamada, David F Wozniak, David M Holtzman,
David H Gutmann, and Michael Wong. Abnormal glutamate homeostasis and
impaired synaptic plasticity and learning in a mouse model of tuberous sclerosis
complex. Neurobiology of Disease, 28(2):184-196, November 2007. 55
[193] Ling-Hui Zeng, Nicholas R Rensing, Bo Zhang, David H Gutmann, Michael J
Gambello, and Michael Wong. Tsc2 gene inactivation causes a more severe
epilepsy phenotype than Tscl inactivation in a mouse model of tuberous sclerosis complex. Hum Mol Genet, 20(3):445-454, February 2011. 57
139